Dissertations / Theses on the topic 'Transfert de style'

L'un des aspects d'une interface homme-machine réussie (p. ex. interaction homme-robot, chatbots, parole, écriture manuscrite, etc.) est la possibilité d'avoir une interaction personnalisée. Cela affecte l'expérience humaine globale et permet une interaction plus fluide. Actuellement, il y a beaucoup de travaux qui utilisent l'apprentissage machine afin de modéliser de telles interactions. Cependant, ces modèles n'abordent pas la question du comportement personnalisé : ils tentent de faire la moyenne des différents exemples provenant de différentes personnes. L'identification des styles humains (persona) ouvre la possibilité de biaiser la sortie des modèles pour prendre en compte la préférence humaine. Dans cette thèse, nous nous sommes concentrés sur le problème des styles dans le contexte de l'écriture manuscrite.L'objectif de cette thèse est d'étudier ces problèmes de styles, dans le domaine de l'écriture. Nous disposons d'un jeu de données IRONOFF, un jeu de données d'écriture manuscrite en ligne, avec 410 rédacteurs, avec ~25K exemples de dessins en majuscules, minuscules et chiffres. Pour le problème de l'apprentissage par transfert, nous avons utilisé un jeu de données supplémentaire, QuickDraw ! (disponible gratuitement sur Google), un jeu de données de dessin d'esquisses contenant environ 50 millions de dessins sur 345 catégories.Les principales contributions de ma thèse sont :1) Proposer un pipeline de travail pour étudier le problème des styles d'écriture. Il s'agit de proposer une méthodologie, des repères et des paramètres d'évaluation (et de fonder ces paramètres d'évaluation).Nous choisissons le paradigme des modèles génératifs temporels dans l'apprentissage profond afin de générer des dessins et d'évaluer leur proximité/pertinence par rapport aux dessins de vérité voulus/de terrain. Nous avons proposé deux métriques, pour évaluer la courbure et la longueur des dessins générés. Afin d'enraciner ces métis, nous avons proposé de multiples repères - dont nous connaissons le pouvoir relatif à l'avance -, puis vérifié que les mesures respectent effectivement la relation de pouvoir relatif.2) Proposer un cadre pour l'étude et l'extraction des styles, et vérifier son avantage par rapport aux repères proposés précédemment.Nous nous sommes mis d'accord sur l'idée d'utiliser un auto-encodeur conditionné en profondeur pour résumer et extraire les informations de style, sans avoir besoin de nous concentrer sur l'identité de la tâche (puisqu'elle est donnée comme une condition). Nous validons ce cadre par rapport au repère proposé précédemment à l'aide de nos paramètres d'évaluation. Nous visualisons également les styles extraits, ce qui nous permet d'obtenir des résultats passionnants !3) En utilisant le cadre proposé, proposer un moyen de transférer l'information sur les styles entre les différentes tâches, et un protocole afin d'évaluer la qualité du transfert.Nous avons exploité le codeur automatique conditionné profond utilisé précédemment, en extrayant la partie codeur - qui, selon nous, contenait les informations pertinentes sur les styles - et en l'utilisant dans de nouveaux modèles formés sur de nouvelles tâches. Nous testons intensivement ce paradigme sur une gamme différente de tâches, à la fois sur les ensembles de données IRONOFF et QuickDraw!. Nous montrons que nous pouvons transférer avec succès les informations de style entre différentes tâches
One aspect of a successful human-machine interface (e.g. human-robot interaction, chatbots, speech, handwriting…,etc) is the ability to have a personalized interaction. This affects the overall human experience, and allow for a more fluent interaction. At the moment, there is a lot of work that uses machine learning in order to model such interactions. However, these models do not address the issue of personalized behavior: they try to average over the different examples from different people in the training set. Identifying the human styles (persona) opens the possibility of biasing the models output to take into account the human preference. In this thesis, we focused on the problem of styles in the context of handwriting.Defining and extracting handwriting styles is a challenging problem, since there is no formal definition for those styles (i.e., it is an ill-posed problem). Styles are both social - depending on the writer's training, especially in middle school - and idiosyncratic - depends on the writer's shaping (letter roundness, sharpness…,etc) and force distribution over time. As a consequence, there are no easy/generic metrics to measure the quality of style in a machine behavior.We may want to change the task or adapt to a new person. Collecting data in the human-machine interface domain can be quite expensive and time consuming. Although most of the time the new task has many things in common with the old task, traditional machine learning techniques fail to take advantage of this commonality, leading to a quick degradation in performance. Thus, one of the objectives of my thesis is to study and evaluate the idea of transferring knowledge about the styles between different tasks, within the machine learning paradigm.The objective of my thesis is to study these problems of styles, in the domain of handwriting. Available to us is IRONOFF dataset, an online handwriting datasets, with 410 writers, with ~25K examples of uppercase, lowercase letters and digits drawings. For transfer learning, we used an extra dataset, QuickDraw!, a sketch drawing dataset containing ~50 million drawing over 345 categories.Major contributions of my thesis are:1) Propose a work pipeline to study the problem of styles in handwriting. This involves proposing methodology, benchmarks and evaluation metrics.We choose temporal generative models paradigm in deep learning in order to generate drawings, and evaluate their proximity/relevance to the intended/ground truth drawings. We proposed two metrics, to evaluate the curvature and the length of the generated drawings. In order to ground those metics, we proposed multiple benchmarks - which we know their relative power in advance -, and then verified that the metrics actually respect the relative power relationship.2) Propose a framework to study and extract styles, and verify its advantage against the previously proposed benchmarks.We settled on the idea of using a deep conditioned-autoencoder in order to summarize and extract the style information, without the need to focus on the task identity (since it is given as a condition). We validate this framework to the previously proposed benchmark using our evaluation metrics. We also to visualize on the extracted styles, leading to some exciting outcomes!3) Using the proposed framework, propose a way to transfer the information about styles between different tasks, and a protocol in order to evaluate the quality of transfer.We leveraged the deep conditioned-autoencoder used earlier, by extract the encoder part in it - which we believe had the relevant information about the styles - and use it to in new models trained on new tasks. We extensively test this paradigm over a different range of tasks, on both IRONOFF and QuickDraw! datasets. We show that we can successfully transfer style information between different tasks

Récemment, les méthodes d'apprentissage profond ont permis d'effectuer des transformations du matériel musical basées sur les données (data-driven). L'objet de cette thèse est le transfert de style musical, dont le but est de transférer de manière automatique le style d'un morceau à un autre. Dans la première partie de ce travail, nous nous concentrons sur les méthodes supervisées pour le transfert de style des accompagnements dans une représentation symbolique, visant à transformer un morceau donné en lui générant un nouvel accompagnement. La méthode proposée est basée sur l'apprentissage supervisé de séquence à séquence à l'aide de réseaux de neurones récurrents (RNN) et s'appuie sur une base de données synthétiques parallèle (alignée par paires) générée à cet effet à l'aide d'un logiciel de génération d'accompagnement existant. Nous proposons ainsi un ensemble de mesures objectives pour évaluer la performance sur cette nouvelle tâche et nous montrons que le système réussit à générer un accompagnement dans le style souhaité tout en suivant la structure harmonique de l'entrée. Dans la deuxième partie, nous étudions une question plus fondamentale : le rôle des encodages positionnels (PE) dans la génération de musique à l'aide des Transformers. Nous proposons l'encodage positionnel stochastique (SPE), un nouveau PE capable de coder des positions relatives et compatible avec une classe récemment proposée de Transformers efficaces. Nous démontrons que le SPE permet, mieux que la méthode conventionnelle (le PE absolu), de modéliser des séquences plus longues que celles rencontrées pendant l'entraînement.Enfin, dans la troisième partie, nous passons de la musique symbolique à l'audio et abordons le problème du transfert de timbre. Plus précisément, nous nous intéressons à transférer le timbre d'un enregistrement audio à un autre, tout en préservant le contenu mélodique et harmonique de ce dernier. Nous présentons une nouvelle méthode pour cette tâche, basée sur une extension de l'autoencodeur variationnel quantifié (VQ-VAE), ainsi qu'une stratégie d'apprentissage auto-supervisé conçue pour obtenir des représentations démêlées du timbre et de la hauteur. Comme dans la première partie, nous concevons un ensemble de métriques objectives pour la tâche. Nous montrons que la méthode proposée est capable de surpasser des méthodes existantes
Recently, deep learning methods have enabled transforming musical material in a data-driven manner. The focus of this thesis is on a family of tasks which we refer to as (one-shot) music style transfer, where the goal is to transfer the style of one musical piece or fragment onto another.In the first part of this work, we focus on supervised methods for symbolic music accompaniment style transfer, aiming to transform a given piece by generating a new accompaniment for it in the style of another piece. The method we have developed is based on supervised sequence-to-sequence learning using recurrent neural networks (RNNs) and leverages a synthetic parallel (pairwise aligned) dataset generated for this purpose using existing accompaniment generation software. We propose a set of objective metrics to evaluate the performance on this new task and we show that the system is successful in generating an accompaniment in the desired style while following the harmonic structure of the input.In the second part, we investigate a more basic question: the role of positional encodings (PE) in music generation using Transformers. In particular, we propose stochastic positional encoding (SPE), a novel form of PE capturing relative positions while being compatible with a recently proposed family of efficient Transformers.We demonstrate that SPE allows for better extrapolation beyond the training sequence length than the commonly used absolute PE.Finally, in the third part, we turn from symbolic music to audio and address the problem of timbre transfer. Specifically, we are interested in transferring the timbre of an audio recording of a single musical instrument onto another such recording while preserving the pitch content of the latter. We present a novel method for this task, based on an extension of the vector-quantized variational autoencoder (VQ-VAE), along with a simple self-supervised learning strategy designed to obtain disentangled representations of timbre and pitch. As in the first part, we design a set of objective metrics for the task. We show that the proposed method is able to outperform existing ones

La génération de gestes expressifs permet aux agents conversationnels animés (ACA) d'articuler un discours d'une manière semblable à celle des humains. Le thème central du manuscrit est d'exploiter et contrôler l'expressivité comportementale des ACA en modélisant le comportement multimodal que les humains utilisent pendant la communication. Le but est (1) d’exploiter la prosodie de la parole, la prosodie visuelle et le langage dans le but de synthétiser des comportements expressifs pour les ACA; (2) de contrôler le style des gestes synthétisés de manière à pouvoir les générer avec le style de n'importe quel locuteur. Nous proposons un modèle de synthèse de gestes faciaux à partir du texte et la parole; et entraîné sur le corpus TEDx que nous avons collecté. Nous proposons ZS-MSTM 1.0, une approche permettant de synthétiser des gestes stylisés du haut du corps à partir du contenu du discours d'un locuteur source et correspondant au style de tout locuteur cible. Il est entraîné sur le corpus PATS qui inclut des données multimodales de locuteurs ayant des styles de comportement différents. Il n'est pas limité aux locuteurs de PATS, et génère des gestes dans le style de n'importe quel nouveau locuteur vu ou non par notre modèle, sans entraînement supplémentaire, ce qui rend notre approche «zero-shot». Le style comportemental est modélisé sur les données multimodales des locuteurs - langage, gestes et parole - et indépendamment de l'identité du locuteur. Nous proposons ZS-MSTM 2.0 pour générer des gestes faciaux stylisés en plus des gestes du haut du corps. Ce dernier est entraîné sur une extension de PATS, qui inclut des actes de dialogue et des repères faciaux en 2D
The generation of expressive gestures allows Embodied Conversational Agents (ECA) to articulate the speech intent and content in a human-like fashion. The central theme of the manuscript is to leverage and control the ECAs’ behavioral expressivity by modelling the complex multimodal behavior that humans employ during communication. The driving forces of the Thesis are twofold: (1) to exploit speech prosody, visual prosody and language with the aim of synthesizing expressive and human-like behaviors for ECAs; (2) to control the style of the synthesized gestures such that we can generate them with the style of any speaker. With these motivations in mind, we first propose a semantically aware and speech-driven facial and head gesture synthesis model trained on the TEDx Corpus which we collected. Then we propose ZS-MSTM 1.0, an approach to synthesize stylized upper-body gestures, driven by the content of a source speaker’s speech and corresponding to the style of any target speakers, seen or unseen by our model. It is trained on PATS Corpus which includes multimodal data of speakers having different behavioral style. ZS-MSTM 1.0 is not limited to PATS speakers, and can generate gestures in the style of any newly coming speaker without further training or fine-tuning, rendering our approach zero-shot. Behavioral style is modelled based on multimodal speakers’ data - language, body gestures, and speech - and independent from the speaker’s identity ("ID"). We additionally propose ZS-MSTM 2.0 to generate stylized facial gestures in addition to the upper-body gestures. We train ZS-MSTM 2.0 on PATS Corpus, which we extended to include dialog acts and 2D facial landmarks

Thesis: S.M., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2018.
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
Cataloged from student-submitted PDF version of thesis.
Includes bibliographical references (pages 41-45).
This thesis studies style transfer on the basis of non-parallel text. This is an instance of a broad family of problems including machine translation, decipherment, and attribute modication. The key challenge is to separate the content from style in an unsupervised manner. We assume a shared latent content distribution across different text corpora, and propose a method that leverages refined alignment of latent representations to perform style transfer. The transferred sentences from one style should match example sentences from the other style as a population. To demonstrate the flexibility of the proposed model, we test it on three tasks: sentiment modication, decipherment of word substitution ciphers, and word order recovery. In both automatic and human evaluation our method achieves strong performance.
by Tianxiao Shen.
S.M.

For many years, light fields have been a unique way of capturing a scene. By using a particular set of optics, a light field camera is able to, in a single moment, take images of the same scene from multiple perspectives. These perspectives can be used to calculate the scene geometry and allow for effects not possible with standard photographs, such as refocus and the creation of novel views.Neural style transfer is the process of training a neural network to render photographs in the style of a particular painting or piece of art. This is a simple process for a single photograph, but naively applying style transfer to each view in a light field generates inconsistencies in coloring between views. Because of these inconsistencies, common light field effects break down.We propose a style transfer method for light fields that maintains consistencies between different views of the scene. This is done by using warping techniques based on the depth estimation of the scene. These warped images are then used to compare areas of similarity between views and incorporate differences into the loss function of the style transfer network. Additionally, this is done in a post-training fashion, which removes the need for a light field training set.

This dissertation is focused on trying to use concepts from style transfer and image-to-image translation to address the problem of defogging. Defogging (or dehazing) is the ability to remove fog from an image, restoring it as if the photograph was taken during optimal weather conditions. The task of defogging is of particular interest in many fields, such as surveillance or self driving cars. In this thesis an unpaired approach to defogging is adopted, trying to translate a foggy image to the correspondent clear picture without having pairs of foggy and ground truth haze-free images during training. This approach is particularly significant, due to the difficult of gathering an image collection of exactly the same scenes with and without fog. Many of the models and techniques used in this dissertation already existed in literature, but they are extremely difficult to train, and often it is highly problematic to obtain the desired behavior. Our contribute was a systematic implementative and experimental activity, conducted with the aim of attaining a comprehensive understanding of how these models work, and the role of datasets and training procedures in the final results. We also analyzed metrics and evaluation strategies, in order to seek to assess the quality of the presented model in the most correct and appropriate manner. First, the feasibility of an unpaired approach to defogging was analyzed, using the cycleGAN model. Then, the base model was enhanced with a cycle perceptual loss, inspired by style transfer techniques. Next, the role of the training set was investigated, showing that improving the quality of data is at least as important as the utilization of more powerful models. Finally, our approach is compared with state-of-the art defogging methods, showing that the quality of our results is in line with preexisting approaches, even if our model was trained using unpaired data.

Thesis: M. Eng., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2018.
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
Cataloged from student-submitted PDF version of thesis.
Includes bibliographical references (pages 46-47).
Style transfer is an active area of research growing in popularity in the Natural Language setting. The goal of this thesis is present a comprehensive review of style transfer tasks used to date, analyze these tasks, and delineate important properties and candidate tasks for future methods researchers. Several challenges still exist, including the difficulty of distinguishing between content and style in a sentence. While some state of the art models attempt to overcome this problem, even tasks as simple as sentiment transfer are still non-trivial. Problems of granularity, transferability, and distinguishability have yet to be solved. I provide a comprehensive analysis of the popular sentiment transfer task along with a number of metrics that highlight its shortcomings. Finally, I introduce possible new tasks for consideration, news outlet style transfer and non-parallel error correction, and provide similar analysis for the feasibility of using these tasks as style transfer baselines.
by Nicholas Matthews.
M. Eng.

In this thesis we will use deep learning tools to tackle an interesting and complex problem of image processing called style transfer. Given a content image and a style image as inputs, the aim is to create a new image preserving the global structure of the content image but showing the artistic patterns of the style image. Before the renaissance of Arti�cial Neural Networks, early work in the �field called texture synthesis, only transferred limited and repeatitive geometric patterns of textures. Due to the avaibility of large amounts of data and cheap computational resources in the last decade Convolutional Neural Networks and Graphics Processing Units have been at the core of a paradigm shift in computer vision research. In the seminal work of Neural Style Transfer, Gatys et al. consistently disentangled style and content from different images to combine them in artistic compositions of high perceptual quality. This was done using the image representation derived from Convolutional Neural Networks trained for large-scale object recognition, which make high level image informations explicit. In this thesis, inspired by the work of Li et al., we build an efficient neural style transfer method able to transfer arbitrary styles. Existing optimisation-based methods (Gatys et al.), produce visually pleasing results but are limited because of the time consuming optimisation procedure. More recent feedforward based methods, while enjoying the inference efficiency, are mainly limited by inability of generalizing to unseen styles. The key ingredients of our approach are a Convolutional Autoencoder and a pair of feature transforms, Whitening and Coloring, reflecting a direct matching of feature covariance of the content image to the given style image. The algorithm allows us to produce images of high perceptual quality that combine the content of an arbitrary photograph with the appearance of arbitrary well known artworks.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2015.
Cataloged from PDF version of thesis.
Includes bibliographical references (pages 139-154).
After taking pictures, photographers often seek to convey their unique moods by altering the style of their photographs, which can involve meticulous contrast management, lighting, dodging, and burning. In this sense, not only are advanced photographers concerned about their pictures' styles; casual photographers who take pictures with cellphone cameras also process their pictures using built-in applications to adjust the image's luminance, coloring, and details. In general, photographers who stylize pictures give them new, different visual appearances, while also preserving the original content. In this context, we investigate problems with novel image stylization, including reproducing the precise time-of-day where the lighting and atmosphere can make a landscape glow, and making a portrait style resemble that created by a renowned photographer. Given an already captured image, however, automatically achieving given styles is challenging. In fact, changing the appearance in a photograph to mimic another time-of-day requires the analysis and modeling of complex 3-D physical light interactions in the scene, while reproducing a portrait photographer's unique style require computers to acquire artistic tastes and a glimpse of the artist's creative process. In this dissertation, we sidestep these Al-complete problems to instead leverage the power of data. We exploit an image database consisting of time-lapse data describing variations in scene appearance during the course of an entire day, and stylish portraits that are already deliberately processed by artists. To leverage these data, we present new algorithms that put input images in dense and local correspondence with examples. In our first method, we change the time-of-day with a single image as the input, which we put in correspondence with a reference time-lapse video. We then extract the local appearance transformations between different frames of the reference, and apply them to the input. In our second method, we transfer the style of a portrait onto a new input by way of local and multi-scale transformations. We demonstrate our methods on public datasets and a large set of photos downloaded from the Internet. We show that we can successfully handle lightings at different times of day and styles by a variety of different artists.
by YiChang Shih.
Ph. D.

This thesis was done in collaboration with Ericsson AB with the goal of researching the possibility of creating a machine learning model that can transfer the style of a text into another arbitrary style depending on the data used. This had the purpose of making their technical documentation appear to have been written with one cohesive style for a better reading experience. Two approaches to solve this task were tested, the first one was to implement an encoder-decoder model from scratch, and the second was to use the pre-trained GPT-2 model created by a team from OpenAI and fine-tune the model on the specific task. Both of these models were trained on data provided by Ericsson, sentences were extracted from their documentation. To evaluate the models training loss, test sentences, and BLEU scores were used and these were compared to each other and with other state-of-the-art models. The models did not succeed in transforming text into a general technical documentation style but a good understanding of what would need to be improved and adjusted to improve the results were obtained.

This thesis was presented on June 22, 2021, the presentation was done online on Microsoft teams. 

Texture transfer of images, or transferring the style of one image to another, has remained one of the challenging problems in the field of image processing. However, a recent algorithmic breakthrough by three German scientists, Gatys, Ecker, and Bethge, which utilizes the technique of neural networks, proved to be quite successful and has generated visually pleasing results. This thesis closely examines the approach to the problem and the architecture of the neural network that these three scientists proposed, provides an implementation of the neural algorithm utilized to generate synthesized images, and presents the subsequent results and analysis.

Thesis (S.M.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2005.
Includes bibliographical references (p. 96-103).
We show that image statistics can be used to analyze and transfer simple notions of pictorial style of paintings and photographs. We characterize the frequency content of pictorial styles, such as multi-scale, spatial variations, and anisotropy properties, using a multi-scale and oriented decomposition, the steerable pyramid. We show that the average of the absolute steerable coefficients as a function of scale characterizes simple notions of "look" or style. We extend this approach to account for image non-stationarity, that is, we capture and transfer the spatial variations of multi-scale content. In addition, we measure the standard deviation of the steerable coefficients across orientation, which characterizes image anisotropy and permits analysis and transfer of oriented structures. We focus on the statistical features that can be transferred. Since we couple analysis and transfer, our statistical model and transfer tools are consistent with the visual effect of pictorial styles. For this reason, our technique leads to more intuitive manipulation and interpolation of pictorial styles. In addition, our statistical model can be used to classify and retrieve images by style.
by Soonmin Bae.
S.M.

Advances within computer vision are being propelled forward by deep convolutional networks and large publicly available datasets. However, for the task of object classification, the accessible datasets usually consist of natural images. Thus, when training networks for classifying objects in paintings, in-domain data is usually limited and natural images are frequently used as a substitute. This often results in lower classification scores due to the domain shift between the training data and the target data domains. In this thesis, we investigated the potential of using the neural artistic style transfer algorithm to create synthetic paintings in order to increase the amount of training data. We collected a dataset of over 14,500 paintings for the purpose of the experiments. A pretrained convolutional neural network was used to extract features from the images. A simple classifier was then trained on a training set of paintings to which different percentages of synthetic images were added. For comparison, the same experiment wasrepeated when natural images were added.We found that adding either synthetic images or natural images to the training data can improve object classification in paintings. Although our results do not support a general preference for synthetic images over natural images, we show that under certain conditions, i.e. for certain classes, adding synthetic images to the training data generated the largest improvement to classification. Finally, we also discuss two main obstacles in using the artistic style transfer algorithm for synthesis of paintings. The first obstacle is the transfer of noise to the synthetic images. The second obstacle is the algorithm’s failure to address aspects of depiction such as the object’s size, the image’s composition, and the change in the object’s appearance over time. This was considered a relevant issue in this thesis as we found that the natural images and the paintings notably differed in these aspects of depiction and in their extracted high-level features.
Djupa faltningsnätverk och stora, offentligt tillgängliga dataset bidrar till stora framsteg inom datorseende. För klassificering av objekt består dom tillgängliga datamängderna oftast av fotografier. När man tränar nätverk för att klassificera objekt i målningar är tillgången till data från samma domän därför oftast begränsad och fotografier används vanligen istället. Detta resulterar dock ofta i ett lägre klassificeringsresultat på grund av domänskiftet mellan träningsdata och måldata.I detta examensarbete undersöker vi möjligheten att använda den neurala stilöverföringsalgoritmen för att skapa syntetiska målningar som kan användas för att öka mängden av träningsdata. Inför våra experiment sammanställde vi ett dataset av över 14 500 målningar. Särdragsrepresentationer extraherades från bilderna med hjälp av ett förtränat faltningsnätverk. En enkel klassificerare tränades sedan på en mängd träningsdata bestående av målningar, till vilket olika procent av syntetiska målningar adderades.Samma experiment upprepades när fotografier adderades.Vi fann att resultatet av klassificeringen kan förbättras genom att öka mängden träningsdata med hjälp av antingen syntetiska målningar eller fotografier. Vårt resultat styrker inte en generell preferens för syntetiska målningar över fotografier. Dock visar vi att under vissa förutsättningar, så som för vissa objektklasser, ger användandet av syntetiska målningar den största förbättringen i klassificering. Slutligen identifierar vi två huvudsakliga problemområden då den neurala stilöverföringsalgoritmen används för att skapa syntetiska målningar för ändamålet att öka mängden av träningsdata. Det första problemet är överföringen av brus till det syntetiska målningarna. Det andra problemet är att algoritmen endast överför målningens stil och därför ignorerar andra skillnader mellan fotografier och målningar i hur objekten skildras så som objektets storlek, bildens sammansättning och hur objektets utseende ändras över tid. Detta ansågs vara ett relevant problem då vi fann att fotografierna och målningarna skiljde sig åt i dessa avseenden samt i deras extraherade särdragsrepresentationer.

It is hoped that the scientific reasoning skills taught in our biology courses will carry over to be applied in novel settings: to new concepts, future courses, other disciplines, and non-academic pursuits. This is the educational concept of transfer. Efforts over many years in the Cell Biology course at BYU to design effective assessment questions that measure competence in both deep understanding of conceptual principles and the ability to draw valid conclusions from experimental data have had at least one disquieting result. The transfer performance of many otherwise capable students is not very satisfactory. In order to explain this unsatisfactory performance, we assumed that the prompts for our transfer problems might be at fault. Consequently, we experimented with multiple versions that differed in wording or the biological setting in which the concept was placed. Performance on the various versions did not change significantly. We are led to investigate two potential underlying causes for this problem. First, like any other important scholastic trait, the ability to transfer requires directed practice through multiple iterations, a feature absent from most courses. Second, perhaps there is something innate about an individual's learning style that is contrary to performing well at transfer tasks. Students sometimes see exams as tests of gamesmanship; "Teachers are trying to outsmart me with trick questions." Post-exam conversations can be very litigious: "But it's not clear what you wanted!" We recommend the pedagogical use of transfer problems which place on the learner the responsibility to define the appropriate scope for inquiry and improve one's ability to acquire the kind of precise and comprehensive understanding that makes transfer possible. In this study, we analyze the effects of directed practice and learning style on transfer abilities. Implications for teaching are discussed and include promoting meta-cognitive practices, carefully selecting lecture and textual materials to reduce the "spotlighting effect" (selective focus on only a subset of ideas), and encouraging students to consciously use multiple learning strategies to help them succeed on various tasks. It is important to note that these skills are likely to take a significant amount of time for both students and teachers to master.

The generation of data has traditionally been specified using hand-crafted algorithms. However, oftentimes the exact generative process is unknown while only a limited number of samples are observed. One such case is generating images that look visually similar to an exemplar image or as if coming from a distribution of images. We look into learning the generating process by constructing a similarity function that measures how close the generated image is to the target image. We discuss a framework in which the similarity function is specified by a pre-trained neural network without fine-tuning, as is the case for neural texture synthesis, and a framework where the similarity function is learned along with the generative process in an adversarial setting, as is the case for generative adversarial networks. The main point of discussion is the combined use of neural networks and maximum mean discrepancy as a versatile similarity function. Additionally, we describe an improvement to state-of-the-art style transfer that allows faster computations while maintaining generality of the generating process. The proposed objective has desirable properties such as a simpler optimization landscape, intuitive parameter tuning, and consistent frame- by-frame performance on video. We use 80,000 natural images and 80,000 paintings to train a procedure for artistic style transfer that is efficient but also allows arbitrary content and style images.
Science, Faculty of
Computer Science, Department of
Graduate

This study investigated the relationship between teacher training transfer, perceptions of principal leadership, and specific principal behaviors among 62 teachers who participated in one district's new teacher induction program. A mixed methods design was employed. A sub-group of 19 participants reported statistically significantly greater training transfer than the other 40 participants; chi-square tests revealed no significant differences between these two groups in terms of demographics or conditions of their supervision. The 19 high-training-transfer participants also reported that their principals demonstrated certain behaviors at a statistically significantly greater level than was reported by the other 40 participants. These principal behaviors were categorized as promoting a culture of accountability and promoting a culture of professional learning. The participants in the high-training-transfer group also were more likely to perceive their principals as strong in both transactional and transformational leadership. However, in the entire sample, leadership style was not found to be significantly related to teacher training transfer, nor were any significant interaction effects revealed between perceived leadership style, teacher demographic variables, and teacher training transfer. The results indicated that specific instructional leadership behaviors were effective in promoting teacher training transfer, and that leadership style made a difference only when applied in the context of these instructional leadership behaviors.

Organisational learning is seen by many to be a key determinant of organisational performance. This is demonstrated by the growth of the 'learning company' concept (Pedler et al. 1991), and by the suggestion that the ability to learn faster than one's competitors is the only sustainable competitive advantage (DeGeus 1988). Consequently, organisations need to integrate and maximise the knowledge and learning of their individuals, and central to the learning process in firms is an effective means of transferring knowledge and learning between individuals and their organisation as a whole. Mental models (individual and shared) have been postulated as a mechanism through which this occurs (Senge 1990a; Kim 1993; Hayes and Allinson 1998). An individual mental model can be characterised as a simplification or representation of understanding of an idea, notion, process or system which provides the cognitive framework in which that individual's knowledge in respect of that issue is stored, whilst shared (group or organisational) mental models can be characterised as the common elements that exist between individual mental models. Both of these have been theoretically linked with individual and organisational learning. Literature in respect of individual and organisational learning, mental models and a third issue cognitive style is reviewed. Cognitive style represents the way individuals obtain, store and operationalise knowledge, and is included here as it is recognised as potentially affecting how learning and mental models interact (Hayes and Allinson 1998). A research model is posited which integrates key theory in respect of these three concepts, and research undertaken in two phases is presented. Phase One focused upon the representation of individual and shared mental models through semi-structured causal interviews with senior mangers in participant organisations, whilst Phase Two involved organisation wide surveys of these models, aspects of learning and cognitive style. Results obtained suggest that the complexities of an organisation, its environment, learning and mental models all mitigate against the identification of a simple relationship between these constructs. However some of the sources of these complexities are identified and suggested, and it is posited that the progression of work addressing organisational learning would best be served through a case study approach addressing the sources of complexity and effectiveness of learning in relation to specific mental models and within organisations.

Les séquences d'ADN peuvent être considérées comme des textes écrits dans un alphabet de 4 lettres. Des techniques inspirées de l'analyse textuelle permettent donc de les caractériser, entre autres à partir de fréquences d'apparition de courtes suites de caractères (les oligonucléotides ou mots). L'ensemble des fréquences des mots d'une longueur donnée est appelé « signature génomique » (cet ensemble est spécifique de l'espèce, ce qui justifie le terme de « signature »). La signature d'espèce est observable sur la plupart des courts fragments d'ADN, ce qui donne à penser qu'elle résulte d'un « style d'écriture ». De plus, la proximité entre espèces du point de vue de la signature génomique correspond bien souvent à une proximité en terme taxonomique. Pourtant, l'analyse des signatures génomiques se confronte rapidement à des limitations dues à la malédiction de la dimension. En effet, les données de grande dimension (la signature génomique a généralement 256 dimensions) montrent des propriétés qui mettent en défaut l'intuition. Par exemple, le phénomène de concentration des distances euclidiennes est bien connu.
Partant de ces constatations, nous avons mis en place des procédures d'évaluation des distances entre signatures de façon à rendre plus manifeste les informations biologiques sur lesquelles s'appuient nos analyses. Une méthode de projection non-linéaire des voisinages y est associée ce qui permet de s'affranchir des problèmes de grande dimension et de visualiser l'espace occupé par les données. L'analyse des relations entre les signatures pose le problème de la contribution de chaque variable (les mots) à la distance entre les signatures. Un Z-score original basé sur la variation de la fréquence des mots le long des génomes a permis de quantifier ces contributions. L'étude des variations de l'ensemble des fréquences le long d'un génomes permet d'extraire des segments originaux. Une méthode basée sur l'analyse du signal permet d'ailleurs de segmenter précisément ces zones originales.
Grâce à cet ensemble de méthodes, nous proposons des résultats biologiques. En particulier, nous mettons en évidence une organisation de l'espace des signatures génomiques cohérente avec la taxonomie des espèces. De plus, nous constatons la présence d'une syntaxe de l'ADN : il existe des « mots à caractère syntaxique » et des « mots à caractère sémantique », la signature s'appuyant surtout sur les mots à caractère syntaxique. Enfin, l'analyse des signatures le long du génome permet une détection et une segmentation précise des ARN et de probables transferts horizontaux. Une convergence du style des transferts horizontaux vers la signature de l'hôte a d'ailleurs pu être observée.
Des résultats variés ont été obtenus par analyse des signatures. Ainsi, la simplicité d'utilisation et la rapidité de l'analyse des séquences par signatures en font un outil puissant pour extraire de l'information biologique à partir des génomes.

When learning a second language, there are elements of a learner's native language that can transfer and are exhibited during production in the second language. This can extend not only to the way things are said but even to gestures that are language- and speech-act-specific. However, there is evidence that the same can occur backwards, that is to say that elements of a second language can be exhibited during production of one's native language (Pavlenko and Jarvis, 2002). This study focuses on English L1 learners of Japanese who have spent significant time both in country and learning the language to see if they exhibit Japanese tendencies when performing apologies in their native English. Comparisons between those with no Japanese experience were made with those who had extensive Japanese experience. Through video recordings of 45 participants engaging in six apology-induced scenarios (non-Japanese, n=24; Japanese, n=21), the participants showed that backward transfer occurs with repetition of IFIDs and nonverbal cues. Further research through different methods can be more telling.

The ability of an organisation to transfer knowledge is one of the key sources of competitive advantage for many of today’s organisations (Argote, 2000). New knowledge is created through interactions between explicit and tacit knowledge (Nonaka and Takeuchi, 1995). From the distinction between explicit and tacit knowledge made by Polanyi’s (1966), it is clear that the former can be transferred with relative ease, particularly using recent advances in information technology. Transfer of tacit knowledge on the other hand, requires social interactions with peers, colleagues, mentors and supervisor (Lahti et al, 2002; Cavusgil et al, 2003). Difficulties associated with this have been referred to as ‘internal stickiness’ (Szulanski, 1996) and is believed to be due to several factors. This study examines difficulties associated with the transfer of managerial tacit knowledge in the relationships involving supervisor and subordinates who work as managers in the Malaysian public sector. After examining previous literature in the field it is hypothesised that the stickiness of knowledge transfer may be associated with the quality of leader member exchange relationships, especially between leaders and their ‘in-group’ versus ‘out-group’ members. For example, in-group relationships are associated with higher levels of trust, respect and obligation compared with out-group relations. Another construct known to be associated with the quality of dyadic relationships is cognitive style (Armstrong, 1999). Cognitive style refers to individual differences in ways of perceiving, organising and processing information and differences in ways in which individuals solve problems, take decisions and relate to others. The research employed a quantitative approach using survey methods. Instruments used in the study included a measure of knowledge transfer stickiness (Szulanski, 1996), Leader Member Exchange (LMX7) (Graen and Uhl-Bien, 1995), Tacit Knowledge Inventory for Managers (TKIM) (Wagner and Sternberg, 1989) and the Cognitive Style Index (CSI) (Allinson and Hayes, 1996). The survey was administered to 1200 managers in the Malaysian Public Sector and 344 completed surveys were returned representing a response rate of 28.7%. Results from a final sample size of 300 managers comprising supervisors and their immediate subordinates are reported. The study successfully determined the relationship between knowledge transfer stickiness, LMX, cognitive style and managerial tacit knowledge. As expected, high-quality LMX leads to higher quality exchanges and concomitant improvements in the transfer of managerial tacit knowledge. Moreover, as hypothesised, individual differences and similarities in cognitive style also influence the transfer of tacit knowledge between supervisor and subordinate. Practical implications are given and recommendation made for future research.

This study sought to understand using qualitative methods why and how the Government of Singapore came to set-up organized research units (ORU) in her two autonomous universities - the National University of Singapore (NUS) and Nanyang Technological University (NTU). That is, are ORUs the chief means in transforming NUS and NTU into top WCUs? The underpinnings of the institutional theory in the globalized context, the theory of academic capitalism, and guidance by frameworks on educational policy transfer, as well as the glonacal heuristics, together with document and discourse analysis of published documents, aided in the design of this study.This study finds that key state and institutional actors in Singapore decided to emulate the key features of U.S.-originated organized research units to make R&D and innovation-led economic growth the vehicle of sustaining this global city-state's global competitiveness in the knowledge-economy and to better elevate the status of both NUS and NTU as world-class. Because global rankings have branding implications, catch-up nations and institutions should make serious attempts to balance the ranking-enhancement effort with their local, national, and regional science-research needs given the constraints of available capital and resources.

Share price periodicity and calendar effects have been well documented for stock exchanges. If these market anomalies are persistent and of sufficiently high amplitudes, the use of frequency analysis will allow investors to earn abnormal returns. This research study examined the use of the discrete Fourier transform combined with prior exponential growth and momentum periodicity as an investment style. A graphical time series approach was used to evaluate performance of the examined styles. The time series consisted of the JSE top 160 shares from December 1985 to October 2013. A momentum-Fourier transform investment style is identified that outperforms most if not all documented univariate ranked investment styles on the JSE for the analysed timeframe. Returns of 27.6% per annum are achieved. It is found that both examined momentum styles are enhanced by using the Fourier transform as a noise filter. Combining prior exponential growth rate and the Fourier transform failed to produce favourable results.
Dissertation (MBA)--University of Pretoria, 2013.
mngibs2014
Gordon Institute of Business Science (GIBS)
MBA
Unrestricted

In Sweden, women of age between of 40 and 74 go through regular screening of their breasts every 18-24 months. The screening mainly involves obtaining a mammogram and having radiologists analyze them to detect any sign of breast cancer. However reading a mammography image requires experienced radiologist, and the lack of radiologist reduces the hospital's operating efficiency. What's more, mammography from different facilities increases the difficulty of diagnosis. Our work proposed a deep learning segmentation system which could adapt to mammography from various facilities and locate the position of the tumor. We train and test our method on two public mammography datasets and do several experiments to find the best parameter setting for our system. The test segmentation results suggest that our system could play as an auxiliary diagnosis tool for breast cancer diagnosis and improves diagnostic accuracy and efficiency.
I Sverige går kvinnor i åldrarna mellan 40 och 74 igenom regelbunden screening av sina bröst med 18-24 månaders mellanrum. Screeningen innbär huvudsakligen att ta mammogram och att låta radiologer analysera dem för att upptäcka tecken på bröstcancer. Emellertid krävs det en erfaren radiolog för att tyda en mammografibild, och bristen på radiologer reducerar sjukhusets operativa effektivitet. Dessutom, att mammografin kommer från olika anläggningar ökar svårigheten att diagnostisera. Vårt arbete föreslår ett djuplärande segmenteringssystem som kan anpassa sig till mammografi från olika anläggningar och lokalisera tumörens position. Vi tränar och testar vår metod på två offentliga mammografidataset och gör flera experiment för att hitta den bästa parameterinställningen för vårt system. Testsegmenteringsresultaten tyder på att vårt system kan fungera som ett hjälpdiagnosverktyg vid diagnos av bröstcancer och förbättra diagnostisk noggrannhet och effektivitet.

Despite the volume, growth, and accessibility of documented knowledge – the insights and experiences stored on paper and in electronic form - management research has yet to demonstrate the same usefulness for documented knowledge as for knowledge residing in human sources. This dissertation explores two areas of potential for documented knowledge, suggesting the efficacy of a piece of documented knowledge is contingent not only on content, but upon the style and structure associated with that content. Style, how cognitively 'concrete' and affectively 'memorable' documented knowledge is perceived to be, is hypothesized to affect how much attention it draws and, in turn, to impact its transfer to users. Structure, reflecting the level of parsimony and modularity in documented knowledge, is hypothesized to impact attention to and manipulation of knowledge such that it affects knowledge transfer and creation. Hypotheses were tested in two laboratory studies using scientific research as an exemplar of documented knowledge. Results indicated that style was associated with documented knowledge, but was not related to its transfer. Likewise, structuring documented knowledge for greater parsimony and modularity did not improve knowledge transfer or knowledge creation. Shortcomings of the empirical tests are evaluated and possibilities for future improvements are discussed.
Ph.D.
Doctorate
Dean's Office, Business Administration
Business Administration
Business Administration; Management

Background: Generative Adversarial Networks (Goodfellow et al., 2014) (GANs)are the current state of the art machine learning data generating systems. Designed with two neural networks in the initial architecture proposal, generator and discriminator. These neural networks compete in a zero-sum game technique, to generate data having realistic properties inseparable to that of original datasets. GANs have interesting applications in various domains like Image synthesis, 3D object generation in gaming industry, fake music generation(Dong et al.), text to image synthesis and many more. Despite having a widespread application domains, GANs are popular for image data synthesis. Various architectures have been developed for image synthesis evolving from fuzzy images of digits to photorealistic images. Objectives: In this research work, we study various literature on different GAN architectures. To understand significant works done essentially to improve the GAN architectures. The primary objective of this research work is synthesis of plant images using Style GAN (Karras, Laine and Aila, 2018) variant of GAN using style transfer. The research also focuses on identifying various machine learning performance evaluation metrics that can be used to measure Style GAN model for the generated image datasets. Methods: A mixed method approach is used in this research. We review various literature work on GANs and elaborate in detail how each GAN networks are designed and how they evolved over the base architecture. We then study the style GAN (Karras, Laine and Aila, 2018a) design details. We then study related literature works on GAN model performance evaluation and measure the quality of generated image datasets. We conduct an experiment to implement the Style based GAN on leaf dataset(Kumar et al., 2012) to generate leaf images that are similar to the ground truth. We describe in detail various steps in the experiment like data collection, preprocessing, training and configuration. Also, we evaluate the performance of Style GAN training model on the leaf dataset. Results: We present the results of literature review and the conducted experiment to address the research questions. We review and elaborate various GAN architecture and their key contributions. We also review numerous qualitative and quantitative evaluation metrics to measure the performance of a GAN architecture. We then present the generated synthetic data samples from the Style based GAN learning model at various training GPU hours and the latest synthetic data sample after training for around ~8 GPU days on leafsnap dataset (Kumar et al., 2012). The results we present have a decent quality to expand the dataset for most of the tested samples. We then visualize the model performance by tensorboard graphs and an overall computational graph for the learning model. We calculate the Fréchet Inception Distance score for our leaf Style GAN and is observed to be 26.4268 (the lower the better). Conclusion: We conclude the research work with an overall review of sections in the paper. The generated fake samples are much similar to the input ground truth and appear to be convincingly realistic for a human visual judgement. However, the calculated FID score to measure the performance of the leaf StyleGAN accumulates a large value compared to that of Style GANs original celebrity HD faces image data set. We attempted to analyze the reasons for this large score.

Formality Style Transfer (FST) is the task of automatically transforming a piece of text from one level of formality to another. Previous research has investigated different methods of performing FST on text in English, but at the time of this project there were to the author’s knowledge no previous studies analysing the quality of FST on text in Swedish. The purpose of this thesis was to investigate how a model trained for FST in Swedish performs. This was done by comparing the quality of a model trained on text in Swedish for FST, to an equivalent model trained on text in English for FST. Both models were implemented as encoder-decoder architectures, warm-started using two pre-existing Bidirectional Encoder Representations from Transformers (BERT) models, pre-trained on Swedish and English text respectively. The two FST models were fine-tuned for both the informal to formal task as well as the formal to informal task, using the Grammarly’s Yahoo Answers Formality Corpus (GYAFC). The Swedish version of GYAFC was created through automatic machine translation of the original English version. The Swedish corpus was then evaluated on the three criteria meaning preservation, formality preservation and fluency preservation. The results of the study indicated that the Swedish model had the capacity to match the quality of the English model but was held back by the inferior quality of the Swedish corpus. The study also highlighted the need for task specific corpus in Swedish.
Överföring av formalitetsstil syftar på uppgiften att automatiskt omvandla ett stycke text från en nivå av formalitet till en annan. Tidigare forskning har undersökt olika metoder för att utföra uppgiften på engelsk text men vid tiden för detta projekt fanns det enligt författarens vetskap inga tidigare studier som analyserat kvaliteten för överföring av formalitetsstil på svensk text. Syftet med detta arbete var att undersöka hur en modell tränad för överföring av formalitetsstil på svensk text presterar. Detta gjordes genom att jämföra kvaliteten på en modell tränad för överföring av formalitetsstil på svensk text, med en motsvarande modell tränad på engelsk text. Båda modellerna implementerades som kodnings-avkodningsmodeller, vars vikter initierats med hjälp av två befintliga Bidirectional Encoder Representations from Transformers (BERT)-modeller, förtränade på svensk respektive engelsk text. De två modellerna finjusterades för omvandling både från informell stil till formell och från formell stil till informell. Under finjusteringen användes en svensk och en engelsk version av korpusen Grammarly’s Yahoo Answers Formality Corpus (GYAFC). Den svenska versionen av GYAFC skapades genom automatisk maskinöversättning av den ursprungliga engelska versionen. Den svenska korpusen utvärderades sedan med hjälp av de tre kriterierna betydelse-bevarande, formalitets-bevarande och flödes-bevarande. Resultaten från studien indikerade att den svenska modellen hade kapaciteten att matcha kvaliteten på den engelska modellen men hölls tillbaka av den svenska korpusens sämre kvalitet. Studien underströk också behovet av uppgiftsspecifika korpusar på svenska.

Thesis (PhD (Chemical Engineering))--Cape Peninsula University of Technology, 2019
The South African wine industry is constantly facing several challenges which affect the quality of wine, the local/global demand and consequently the revenue generated. These challenges include the ongoing drought, bush fires, climate change and several liquor amendment bills aimed at reducing alcohol consumption and alcohol outlets in South Africa. It is therefore critical for the wine industry to expand and find alternative ways in which sub-standard or surplus wine grapes can be used to prevent income losses and increase employment opportunities. Traditional Balsamic Vinegar (TBV) is a geographically and legislative protected product produced only in a small region in Italy. However, the methodology can be used to produce similar vinegars in other regions. Balsamic-styled vinegar (BSV), as defined in this thesis, is a vinegar produced by partially following the methods of TBV while applying process augmentation techniques. Balsamic-styled vinegar is proposed to be a suitable product of sub-standard quality or surplus wine grapes in South Africa. However, the production of BSV necessitates the use of cooked (high sugar) grape must which is a less favourable environment to the microorganisms used during fermentation. Factors that negatively affect the survival of the microorganisms include low water activity due to the cooking, high osmotic pressure and high acidity. To counteract these effects, methods to improve the survival of the non-Saccharomyces yeasts and acetic acid bacteria used are essential. The primary aim of this study was to investigate several BSV process augmentation techniques such as, aeration, agitation, cell immobilization, immobilized cell reusability and oxygen mass transfer kinetics in order to improve the performance of the microbial consortium used during BSV production. The work for this study was divided into four (4) phases. For all the phases a microbial consortium consisting of non-Saccharomyces yeasts (n=5) and acetic acid bacteria (n=5) was used. Inoculation of the yeast and bacteria occurred simultaneously. The 1st phase of the study entailed evaluating the effect of cells immobilized by gel entrapment in Ca-alginate beads alongside with free-floating cells (FFC) during the production of BSV. Two Ca-alginate bead sizes were tested i.e. small (4.5 mm) and large (8.5 mm) beads to evaluate the effects of surface area or bead size on the overall acetification rates. Ca-alginate beads and FFC fermentations were also evaluated under static and agitated (135 rpm) conditions. The 2nd phase of the study involved studying the cell adsorption technique for cell immobilization which was carried-out using corncobs (CC) and oak wood chips (OWC), while comparing to FFC fermentations. At this phase of the study, other vinegar bioreactor parameters such as agitation and aeration were studied in contrast to static fermentations. One agitation setting (135 rpm) and two aeration settings were tested i.e. high (0.3 vvm min−1) and low (0.15 vvm min−1) aeration conditions. Furthermore, to assess the variations in cell adsorption capabilities among individual yeast and AAB cells, the quantification of cells adsorbed on CC and OWC prior- and post-fermentation was conducted using the dry cell weight method. The 3rd phase of the study entailed evaluating the reusability abilities of all the matrices (small Ca-alginate beads, CC and OWC) for successive fermentations. The immobilized cells were evaluated for reusability on two cycles of fermentation under static conditions. Furthermore, the matrices used for cell immobilization were further analysed for structure integrity by scanning electron microscopy (SEM) before and after the 1st cycle of fermentations. The 3rd phase of the study also involved the sensorial (aroma and taste) evaluations of the BSV’s obtained from the 1st cycle of fermentation in order to understand the sensorial effects of the Ca-alginate beads, CC and OWC on the final BSV. The 4th phase of the study investigated oxygen mass transfer kinetics during non-aerated and aerated BSV fermentation. The dynamic method was used to generate several dissolved oxygen profiles at different stages of the fermentation. Consequently, the data obtained from the dynamic method was used to compute several oxygen mass transfer parameters, these include oxygen uptake rate ( 𝑟𝑟𝑂𝑂2 ), the stoichiometric coefficient of oxygen consumption vs acid yield (𝑌𝑌𝑂𝑂/𝐴𝐴), the oxygen transfer rate (𝑁𝑁𝑂𝑂2 ), and the volumetric mass transfer coefficients (𝐾𝐾𝐿𝐿𝑎𝑎). During all the phases of the study samples were extracted on weekly intervals to evaluate pH, sugar, salinity, alcohol and total acidity using several analytical instruments. The 4th phase of the study involved additional analytical tools, i.e. an oxygen µsensor to evaluate dissolved oxygen and the ‘Speedy breedy’ to measure the respiratory activity of the microbial consortium used during fermentation. The data obtained from the 1st phase of the study demonstrated that smaller Ca-alginate beads resulted in higher (4.0 g L-1 day−1) acetification rates compared to larger (3.0 g L-1 day−1) beads, while freely suspended cells resulted in the lowest (0.6 g L-1 day−1) acetification rates. The results showed that the surface area of the beads had a substantial impact on the acetification rates when gel entrapped cells were used for BSV fermentation. The 2nd phase results showed high acetification rates (2.7 g L-1 day−1) for cells immobilized on CC in contrast to cells immobilized on OWC and FFC, which resulted in similar and lower acetification rates. Agitated fermentations were unsuccessful for all the treatments (CC, OWC and FFC) studied. Agitation was therefore assumed to have promoted cell shear stress causing insufficient acetification during fermentations. Low aerated fermentations resulted in better acetification rates between 1.45–1.56 g L-1 day−1 for CC, OWC and FFC. At a higher aeration setting, only free-floating cells were able to complete fermentations with an acetification rate of 1.2 g L-1 day−1. Furthermore, the adsorption competence data showed successful adsorption on CC and OWC for both yeasts and AAB with variations in adsorption efficiencies, whereby OWC displayed a lower cell adsorption capability compared to CC. On the other hand, OWC were less efficient adsorbents due to their smooth surface, while the rough surface and porosity of CC led to improved adsorption and, therefore, enhanced acetification rates. The 3rd phase results showed a substantial decline in acetification rates on the 2nd cycle of fermentations when cells immobilized on CC and OWC were reused. While cells entrapped in Ca-alginate beads were able to complete the 2nd cycle of fermentations at reduced acetification rates compared to the 1st cycle of fermentations. The sensory results showed positive ratings for BSV’s produced using cells immobilized in Ca-alginate beads and CC. However, BSV’s produced using OWC treatments were neither ‘liked nor disliked’ by the judges. The SEM imaging results further showed a substantial loss of structural integrity for Ca-alginate beads after the 1st cycle fermentations, with minor changes in structural integrity of CC being observed after the 1st cycle fermentations. OWC displayed the same morphological structure before and after the 1st cycle fermentations which was attributed to their robustness. Although Ca-alginate beads showed a loss in structural integrity, it was still assumed that Ca-alginate beads provided better protection against the harsh environmental conditions in contrast to CC and OWC adsorbents due to the acetification rates obtained on both cycles. The 4th phase data obtained from the computations showed that non-aerated fermentations had a higher 𝑌𝑌𝑂𝑂/𝐴𝐴, 𝑟𝑟𝑂𝑂2 , 𝑁𝑁𝑂𝑂2 and a higher 𝐾𝐾𝐿𝐿𝑎𝑎 . It was clear that aerated fermentations had a lower aeration capacity due to an inappropriate aeration system design and an inappropriate fermentor. Consequently, aeration led to several detrimental biochemical changes in the fermentation medium thus affecting 𝐾𝐾𝐿𝐿𝑎𝑎 and several oxygen mass transfer parameters which serve as a driving force. Overall, it was concluded that the best method for BSV production is the use of cells entrapped in small alginate beads or cells adsorbed on CC under static and non-aerated fermentations. This conclusion was based on several factors such as cell affinity/cell protection, acetification rates, fermentation period and sensorial contributions. However, cells entrapped in Ca-alginate beads had the highest acetification rates. The oxygen mass transfer computations demonstrated a high 𝐾𝐾𝐿𝐿𝑎𝑎 when Ca-alginate beads were used under static-non-aerated conditions compared to fermentations treated with CC. Therefore, a fermentor with a high aeration capacity needs to be designed to best suit the two BSV production systems (Ca-alginate beads and CC). It is also crucial to develop methods which can increase the robustness of Ca-alginate beads in order to improve cell retention and reduce the loss of structural integrity for subsequent cycles of fermentation. Studies to define parameters used for upscaling the BSV production process for large scale productions are also crucial.

This corpus-based study investigates language use in the occluded genre of written pleadings before the European Court of Human Rights through the paradigms of legal phraseology and Translation Studies. The analysis is carried out on three subcorpora of authentic texts: (a) pleadings translated from Russian into English, (b) pleadings translated from Italian into English and (c) pleadings originally drafted in English. Legal language is intricate and formulaic, and frequently makes recourse to prefabricated patterns and routines. Legal phraseology is a major challenge for professional legal translators, and yet its translation has not received much scholarly attention until recently. Legal phraseological units are prefabricated patterns that form the matrix of legal texts and reveal interesting information about both the language and structure of the genre of written pleadings. Over the last thirty years, linguistic deviations occurring in the translation process have constituted one of the main areas within Translation Studies. It has been postulated that translated language has distinctive linguistic characteristics. Legal translation, in addition to linguistic factors, is conditioned by the tension between the legal systems involved, which can result in peculiar language dynamics in the translation of legal texts. This study draws inspiration from Toury’s (1995) and Chesterman’s (2004a) works to describe the different dynamics of translated language, applying a combination of translation norms and universals to identify and describe regularities in translated pleadings. This work is carried out using both linguistic and translational insights in order to demonstrate empirically how written pleadings can be characterised in terms of their phraseological content and how translated pleadings differ from non-translated pleadings. Distributional patterns of recurrent and anomalous legal phraseological units are compared across the corpora and analysed for typicality of frequencies and patterning as well as for quantity and quality of linguistic variation. The results contain a list of legal style markers typical of this genre, obtained in a translational and phraseological perspective. The list supplements the rather scant information about the language of written pleadings at the European Court of Human Rights. The analysis also provides confirmatory evidence of the differences between translated and non-translated texts, specifically, proving the co-existence of two opposite tendencies in translation: conventionalisation (translation of source text textemes with conventional repertoremes of the target environment) and discourse transfer (introduction of prefabricated patterns from the source language). The results may also be of some use and applicability for Russian-to-English and Italian-to-English translators, helping them avoid interference, use of unnatural or overly conservative patterns.

The transformation of economic and societal contexts often labeled sharing economy involves changes in business processes, interactions between people and use of information and communication technologies (ICTs) [...]. Now, platform brands and digital cues court for consumer attention in stylized customer experience journeys. Due to the flexibility and internationalization of actors, processes and organizational structures, research has ‘largely centered on marketplace design, yield management, choosing an appropriate organizational scope, and creating effective labor policy’ [...]. [Aus der Einleitung.]

Software as a service is an emerging but important aspect of the web. WebServices play a vital role in providing it. Web Services are commonly provided in one of two architectural styles: a "REpresentational State Transfer" (REST), or using the "Simple Object Access Protocol" (SOAP.) Originally most web content was text and small images. But more recent services involve complex data structures including text, images, audio, and video. The task of optimizing data to provide delivery of these structures is a complex one, involving both theoretical and practical aspects. In this thesis work, I have considered two architectures developed in the REST architectural style and tested them on mixes of data types (plain text, image, audio) being retrieved from a file system or database. The payload which carries the actual content of a data transmission process can either be in Extensible Markup Language (XML) or JavaScript Object Notation (JSON). Both of these language notations are widely used. The two architectures used in this thesis work are titled as Scenario 1 and Scenario 2. Scenario 1 proposes two different cases for storing, retrieving and presenting the data via a REST web service. We investigate the question of what is the best way to provide different data types (image, audio) via REST Web Service. Payload size for JSON and XML are compared. Scenario 2 proposes an enhanced and optimized architecture which is derived from the pros of the first two cases in Scenario 1. The proposed architecture is best suited for retrieving and serving non-homogeneous data as a service in a homogenous environment. This thesis is composed of theoretical and practical parts. The theory part contains the design and principles of REST architecture. The practical part has a Web Service provider and consumer model developed in Java. The practical part is developed using the Spring MVC framework and Apache CXF, which provides an implementation using JAX-RS, the Java API for RESTful services. A glossary of acronyms used in this thesis appears in the appendix on page 101.

Machine Translation (MT) is the task of mapping a source language to a target language. The recent introduction of neural MT (NMT) has shown promising results for high-resource language, however, poorly performing for low-resource language (LRL) settings. Furthermore, the vast majority of the 7, 000+ languages around the world do not have parallel data, creating a zero-resource language (ZRL) scenario. In this thesis, we present our approach to improving NMT for LRL and ZRL, leveraging a multilingual NMT modeling (M-NMT), an approach that allows building a single NMT to translate across multiple source and target languages. This thesis begins by i) analyzing the effectiveness of M-NMT for LRL and ZRL translation tasks, spanning two NMT modeling architectures (Recurrent and Transformer), ii) presents a self-learning approach for improving the zero-shot translation directions of ZRLs, iii) proposes a dynamic transfer-learning approach from a pre-trained (parent) model to a LRL (child) model by tailoring to the vocabulary entries of the latter, iv) extends M-NMT to translate from a source language to specific language varieties (e.g. dialects), and finally, v) proposes an approach that can control the verbosity of an NMT model output. Our experimental findings show the effectiveness of the proposed approaches in improving NMT of LRLs and ZRLs.

Machine Translation (MT) is the task of mapping a source language to a target language. The recent introduction of neural MT (NMT) has shown promising results for high-resource language, however, poorly performing for low-resource language (LRL) settings. Furthermore, the vast majority of the 7, 000+ languages around the world do not have parallel data, creating a zero-resource language (ZRL) scenario. In this thesis, we present our approach to improving NMT for LRL and ZRL, leveraging a multilingual NMT modeling (M-NMT), an approach that allows building a single NMT to translate across multiple source and target languages. This thesis begins by i) analyzing the effectiveness of M-NMT for LRL and ZRL translation tasks, spanning two NMT modeling architectures (Recurrent and Transformer), ii) presents a self-learning approach for improving the zero-shot translation directions of ZRLs, iii) proposes a dynamic transfer-learning approach from a pre-trained (parent) model to a LRL (child) model by tailoring to the vocabulary entries of the latter, iv) extends M-NMT to translate from a source language to specific language varieties (e.g. dialects), and finally, v) proposes an approach that can control the verbosity of an NMT model output. Our experimental findings show the effectiveness of the proposed approaches in improving NMT of LRLs and ZRLs.

AbstractTitle:Football coaches’ awareness and implementation of team dynamicsObjectives: This research aims to explore, based on the in-depth interviews with Czech male football coaches, how the football coaches work with team dynamics and its principles, during recruitment/transfers, training process, and specific situations, and potentially ascertain what are the obstacles for the team dynamics principles implementation.Methods:The primary method used in this research is the qualitative method of semistructured in-depth interviews. Research is focused on the population of male Czech football coaches with at least half a year of experience. The research sample consists of 7 respondents. The data are proceeded according to GDPR and anonymized. For the data interpretation, the interpretative phenomenological analysis (IPA) is used. Results:The respondents are aware of team cohesion, climate, and synergy and understand the importance of team dynamics in general. If it is possible, the coaches look at the players' psychological characteristics and consider them in relation to the rest of the team and even their potential behavior. The coaches’ behavior and interventions differ based on the situation. However, they tend not to intervene much, and they also don't use almost any analytical tools in terms of team dynamics. That can be caused by many factors such as they perceived the team dynamics as part of psychology, which according to this research, might be perceived as a tool for problem treatment, not as a preventive tool. There are also some structural barriers such as unclear club concept, unclear competency among the responsible, not enough emphasis on team dynamics during the licenses education and players’ agents pressure, as well as language barrier, which seems to prevent successful implementation of foreign players. The respondents tend to rationalize some of the problems. Keywords:cohesion, team climate, synergy, competencies, recruitment, training process, leadership styles, systematical obstacles, barriers

Histopathological evaluation and Gleason grading on Hematoxylin and Eosin(H&E) stained specimens is the clinical standard in grading prostate cancer. Recently, deep learning models have been trained to assist pathologists in detecting prostate cancer. However, these predictions could be improved further regarding variations in morphology, staining and differences across scanners. An approach to tackle such problems is to employ conditional GANs for style transfer. A total of 52 prostatectomies from 48 patients were scanned with two different scanners. Data was split into 40 images for training and 12 images for testing and all images were divided into overlapping 256x256 patches. A segmentation model was trained using images from scanner A, and the model was tested on images from both scanner A and B. Next, GANs were trained to perform style transfer from scanner A to scanner B. The training was performed using unpaired training images and different types of Unsupervised Image to Image Translation GANs (CycleGAN and UNIT). Beside the common CycleGAN architecture, a modified version was also tested, adding Kullback Leibler (KL) divergence in the loss function. Then, the segmentation model was tested on the augmented images from scanner B.The models were evaluated on 2,000 randomly selected patches of 256x256 pixels from 10 prostatectomies. The resulting predictions were evaluated both qualitatively and quantitatively. All proposed methods outperformed in AUC, in the best case the improvement was of 16%. However, only CycleGAN trained on a large dataset demonstrated to be capable to improve the segmentation tool performance, preserving tissue morphology and obtaining higher results in all the evaluation measurements. All the models were analyzed and, finally, the significance of the difference between the segmentation model performance on style transferred images and on untransferred images was assessed, using statistical tests.

The industry of cinema has experienced a radical change in the last decades: the transition from film cinematography to its digital format. As a consequence, several challenges have appeared, but, at the same time, many possibilities are open now for cinematographers to explore with this new medium. In this thesis, we propose different tools that can be useful for cinematographers while doing their craft. First, we develop a tool for automatic color grading. It is a statistics-based method to automatically transfer the style from a graded image to unprocessed footage. Some advantages of the model are its simplicity and low computational cost, which make it amenable for real-time implementation, allowing cinematographers to experiment on-set with different styles and looks. Then, a method for adding texture to footage is created. In cinema, the most commonly used texture is film grain, either directly shooting on film, or adding synthetic grain later-on at post-production stage. We propose a model of "retinal noise" which is inspired by processes in the visual system, and produces results that look natural and visually pleasing. It has parameters that allow to vary widely the resulting texture appearance, which make it an artistic tool for cinematographers. Moreover, due to the "masking" phenomenon of the visual system, the addition of this texture improves the perceived visual quality of images, resulting in bit rate and bandwidth savings. The method has been validated through psychophysical experiments in which observers, including cinema professionals, prefer it over film grain emulation alternatives from academia and the industry. Finally, we introduce a physiology-based image quality metric, which can have several applications in the image processing field, and more specifically in the cinema and broadcasting context: video coding, image compression, etc. We study an optimization of the model parameters in order to be competitive with the state-of-the-art quality metrics. An advantage of the method is its reduced number of parameters, compared with some state-of-the-art methods based in deep-learning, which have a number of parameters several orders of magnitude larger.
La industria del cine ha experimentado un cambio radical en las últimas décadas: la transición de su soporte fílmico a la tecnología del cine digital. Como consecuencia, han aparecido algunos desafíos técnicos, pero, al mismo tiempo, infinitas nuevas posibilidades se han abierto con la utilización de este nuevo medio. En esta tesis, se proponen diferentes herramientas que pueden ser útiles en el contexto del cine. Primero, se ha desarrollado una herramienta para aplicar \textit{color grading} de manera automática. Es un método basado en estadísticas de imágenes, que transfiere el estilo de una imagen de referencia a metraje sin procesar. Las ventajas del método son su sencillez y bajo coste computacional, que lo hacen adecuado para ser implementado a tiempo real, permitiendo que se pueda experimentar con diferentes estilos y 'looks', directamente on-set. En segundo lugar, se ha creado un método para mejorar imágenes mediante la adición de textura. En cine, el grano de película es la textura más utilizada, ya sea porque la grabación se hace directamente sobre película, o porque ha sido añadido a posteriori en contenido grabado en formato digital. En esta tesis se propone un método de 'ruido retiniano' inspirado en procesos del sistema visual, que produce resultados naturales y visualmente agradables. El modelo cuenta con parámetros que permiten variar ampliamente la apariencia de la textura, y por tanto puede ser utilizado como una herramienta artística para cinematografía. Además, debido al fenómeno de enmascaramiento del sistema visual, al añadir esta textura se produce una mejora en la calidad percibida de las imágenes, lo que supone ahorros en ancho de banda y tasa de bits. El método ha sido validado mediante experimentos psicofísicos en los cuales ha sido elegido por encima de otros métodos que emulan grano de película, métodos procedentes de academia como de industria. Finalmente, se describe una métrica de calidad de imágenes, basada en fenómenos fisiológicos, con aplicaciones tanto en el campo del procesamiento de imágenes, como más concretamente en el contexto del cine y la transmisión de imágenes: codificación de vídeo, compresión de imágenes, etc. Se propone la optimización de los parámetros del modelo, de manera que sea competitivo con otros métodos del estado del arte . Una ventaja de este método es su reducido número de parámetros comparado con algunos métodos basados en deep learning, que cuentan con un número varios órdenes de magnitud mayor.

碩士
國立中興大學
資訊科學與工程學系
106
In recent years, artificial intelligence has flourished, and deep learning has been widely used in various research fields and has made significant breakthroughs in various applications such as translation, face recognition, voice assistant, and spam detection. Style transfer is an important research project for deep learning and has achieved remarkable results in the field of computer vision. However, it is relatively immature in the field of natural language processing. One of the main reasons is that the existing text style transfer methods require a large number of parallel corpora for style transfer. In this paper, in order to solve the problem of lack of parallel corpus, we propose a two-stage unsupervised training framework to achieve unsupervised text style transfer tasks. The core idea is to (1) separately train two models from the source style text and the target style text to capture the semantics behind the texts, and (2) combine the trained model to perform style transfer. By doing so, we need only non-parallel text corpus to perform style transfer. In this thesis, we use the modern and wuxia styles as examples to validate the proposed framework and report research findings.

碩士
國立臺灣大學
電信工程學研究所
107
This paper mainly discusses the comparison between style conversion algorithms. Through the feature extraction on the image through the convolution neural network, the image style and structure information can be captured and reconstructed. VGG-NET can help the computer vision field extract a lot of features in the image, and through the operation of the feature to complete the style conversion work, can achieve different style effects through the setting of many parameters, but the distortion in the portrait details is the main problem we found. This greatly limits its application. By gradually processing different intervals, we have completed the situation where the style conversion is prone to distortion of the style avatar. The first part of the paper focuses on the idea that Getys proposed in 2015. The algorithm for style conversion through the curly network, using the different curl layer features in VGG-NET for reconstruction and Gram-matrix design, has achieved good results, and can know the image characteristics of the curl network observed at different levels. Differences, according to different levels of image features, can achieve different conversion effects. This also makes us know that the curly network can get rich and rich features, and also proposes new imagination for style conversion. However, the computational complexity limits its application, so many algorithms are proposed to improve the execution speed. The second part of the thesis mainly uses the coding and decoder design, Whiting & Coloring, feature mapping design, and also uses the different curl layer features in VGG-NET to reconstruct, and obtain different conversion effects. Improve the efficiency of style conversion while completing the operability of any style conversion. The third part is that we find that the portrait style conversion is easy to be distorted, so we cut the image so that the neural network can capture the features of the key image more accurately, so that the style conversion can be used more easily in everyday devices. And operation. Therefore, our proposed algorithm utilizes the current universal style transfer via feature transforms, uses image pre-processing, performs segmentation, and individually completes the reconstructed image after style conversion, so that the style conversion algorithm can also be used in portrait photos.

碩士
國立臺灣大學
資訊工程學研究所
107
Style transfer is a popular topic in computer science. There are lots of works try to turn an image into the desired style. In addition, there are some popular commercial products. With the growth of head-mounted display, the demand of 3D multimedia content is increasing, such as games and virtual reality applications. Due to creating a scene in a particular style is time-consuming, we take style transfer into the design process. However, most of these works focus on 2D images. We modify these methods to reach our goal. In this thesis, we propose a method to transfer the 3D scene style by a specific painting. In the preprocessing stage, we transfer the style of albedo maps. We also generate normal maps and displacement maps according to the transferred albedo maps. In the runtime, we tune the lighting effect and add silhouette to improve the transferred result. Finally, we demonstrate the result of transferring objects by different areas of the style image and discuss the runtime performance.

Dissertação de Mestrado Integrado em Engenharia Electrotécnica e de Computadores apresentada à Faculdade de Ciências e Tecnologia
Transferência de estilo com redes neuronais é a versão mais recente do ramo de interpretações artísticas baseadas em imagens. Historicamente, algoritmos de estilização para interpretações não-realistas foram desenvolvidos especificamente em torno de certas primitivas. Por exemplo, uma interpretação baseada em pinceladas colocava pinceladas virtuais numa imagem, mas era desenvolvida cuidadosamente para um estilo particular de pincelada e revelava-se incapaz de simular um estilo arbitrário. Esta limitação inerente de flexibilidade, estilo e diversidade que alguns algoritmos de interpretações artísticas baseadas em imagens tinham era equilibrada pela sua capacidade de representar fielmente os estilos artísticos para os quais eram criados. A procura por novos algoritmos que respondessem a estas limitações resulta no aparecimento da transferência de estilo com redes neuronais. A introdução de redes neuronais convolucionais causou uma mudança profunda nesta velha área de investigação, e atraiu a atenção de círculos académicos e industriais.Esta dissertação tem como objectivo ultrapassar as limitações computacionais do algoritmo clássico de transferência de estilo treinando uma rede geradora para realizar a mesma tarefa centenas de vezes mais depressa. A continuação lógica de uma transferência de estilo mais rápida, que é a transferência de estilo em vídeo, é um tópico que será explorado nesta dissertação devido às suas variadas aplicações em cenários de realidade aumentada e de realidade virtual, e na indústria de animação. Para solucionar o problema de processamento de vídeo com redes neuronais, duas alternativas são consideradas: utilizar métodos do ramo de visão por computador para guiar o treino da rede, ou alterar a arquitectura da rede para aferir informação temporal e espacial ao mesmo tempo.
Neural style transfer is the most recent facet of image-based artistic rendering. Historically, stylization algorithms for non-photorealistic rendering were designed specifically around certain primitives. For example, stroke based rendering placed virtual strokes on an image, but was carefully designed for only one particular style of stroke and not capable of simulating an arbitrary style. This inherent limitation on flexibility, style and diversity some IB-AR algorithms had was balanced by their capability of faithfully depicting those certain prescribed styles. The demand for novel algorithms to address these limitations gives birth to the field of NST. The introduction of convolutional neural networks caused a paradigm shift in this long standing area of research, and attracted the attention of both academic and industrial circles. This dissertation has the goal of enabling classical neural style transfer to overcome its computational limitations by training a generative network to perform the same task hundreds of times faster. The logical continuation of faster neural style transfer, video style transfer, is a topic that will be explored due to its many possible applications in augmented reality and virtual reality scenarios, and in the animation industry. To solve the task of video processing with neural networks, two alternatives are considered: using computer vision methods to guide a network's training, or changing a networks architecture to take into account spatial and temporal information at the same time.

碩士
國立交通大學
多媒體工程研究所
105
In this thesis, we propose a multi-level shape style transfer method. We first apply mesh parameterization to project input patches to U-V domain meshes with free-boundary, and then perform remeshing to re-sample U-V domain meshes to image space. Each vertex in this mesh stores its original 3D position. We split shape style into 2 degrees, the first is called “base shape”, and the other is called “detail components”. Both degrees of style are computed through Fourier transform in image space. We also prevent aliasing problem during remeshing to image space by computing the detail components through subtraction of original patch by reconstructed base shape in 3D space. The final transfer process allows user adding, subtracting, or blending selected styles and transfer to selected source model.

博士
國立中興大學
資訊科學與工程學系
103
In this study we present two steganographic algorithms using texture synthesis and style transfer. The first algorithm we propose is a novel approach for steganography using a reversible texture synthesis. A texture synthesis process re-samples a smaller texture image which synthesizes a new texture image with a similar local appearance and arbitrary size. We weave the texture synthesis process into steganography to conceal secret messages. In contrast to using an existing cover image to hide messages, our algorithm conceals the source texture image and embeds secret messages through the process of texture synthesis. This allows us to extract secret messages and the source texture from a stego synthetic texture. Our approach offers three distinct advantages. First, our scheme offers the embedding capacity that is proportional to the size of the stego texture image. Second, a steganalytic algorithm is not likely to defeat our steganographic approach. Third, the reversible capability inherited from our scheme provides functionality which allows recovery of the source texture. Experimental results have verified that our proposed algorithm can provide various numbers of embedding capacities, produce a visually plausible texture images, and recover the source texture. The second algorithm we introduce is a steganographic algorithm that using constrained texture synthesis and style transfer. Given an original source texture and a target image or vector field, our scheme can produce a new stego synthetic image concealing secret messages. Our approach is the first steganographic method that uses the constrained texture synthesis and style transfer scheme to convey secret message. We recommend using the combinatorial number system to provide an efficient data embedding solution. Experimental results verified that our proposed algorithm can provide various numbers of embedding capacities, produce a visually plausible texture images, and recover the source texture. To best of our knowledge, these two algorithms are novel and are the first attempt of using the texture synthesis and style transfer techniques to achieve steganography.

碩士
國立交通大學
多媒體工程研究所
107
Neural style transfer is usually suitable for use in abstract styles. When used in styles such as Japanese animation whose foreground is more complex than their background, the results are often not as good as expected. We design a method to automatically transfer the style for video with this type of style. We combine semantic segmentation and spatial control to transfer the specified style to the specified area. By designing the initial image and the loss function, we fixed the distortion of the face and the incomplete style transfer. We propose a method to provide users with the ability to adjust the feature weights of different regions to maintain the artistic conception of the target style, we also combine the optical flow to ensure the coherence from frame to frame in the video.

One of the exciting research field has implemented known as Neural Style Transfer, which is a technique to transform images in an artistic way. Two images are taken as input image namely style image and content image to transform another base image with the help of optimization technique. This NST can be done with the help of Convolutional Neural Networks model as many researchers have tried to achieve good results using CNN network architecture. One of the famous and efficient pre-trained architecture is VGG16 and Gatys et al. [2] were able to generated good results based upon the VGG model. [2] Many famous Mobile and Web applications like DeepArt, Prisma and Pikazoapp have used these models to transformed images in an artistic way. [6] [27] We primarily have discussed different Neural Style Transfer techniques then we have classified the artistic style transfer. We have implementation the model in Keras with the pretrained CNN model that is VGG19 where we have adjusted the hyperparameters and transformation coefficients. VGG19 model has been trained on ImageNet dataset and we used it for feature extraction where for testing we have used two datasets namely Caltech101 and Caltech256. The fundamentals of NST is also discussed in depth literature survey which can be found in chapter 2.

碩士
國立東華大學
資訊工程學系
95
Digital painting synthesizes an output image with paint styles of example images along with the input source image. However, the synthesis procedure always requires the user intervention in selecting patches from example images that best describe its paint styles. The thesis presents a systematic system framework to synthesize example-based rendering images requires no user intervention in the synthesis procedure. The artistic database is been comprised in this work, and the user can synthesize an image according to the paint styles of different well known artists. We use the mean shift image segmentation procedure and the texture re-synthesis method to construct our artistic database, and then find the correspondence between example textures and the mean-shifting areas of the input source image, and then synthesize the output images using the patch-based sampling approach. The main contribution of this thesis is the systematic paint style transfer system for synthesizing a new image without requiring any user intervention. The artistic database is composed of re-synthesized mean-shifting example images of different artists, which are adopted as learning examples of the paint style of different well known artists during the synthesis procedure, and the system will synthesize a new image with the paint style of the user selected artist from the database automatically.

碩士
國立中正大學
資訊工程研究所
104
In the past, visual arts usually represented the static art like paintings, photography and sculptures. In recent years, many museums, artwork galleries, and even art exhibitions demonstrated dynamic artworks for visitors to relish. The most famous dynamic artwork is “The moving painting of Along the River During the Qingming Festival”. Nevertheless, it took two years to complete this work. They had to plan each action for every character at first, then drew each video frame by animators. Finally, it could achieve seamless stitching by using lots of projectors to render scene on the screen. In our research, we propose a method for generating animated paintings. It only needs millions of videos on a network of existing databases and requires users to perform some simple auxiliary operations to achieve the effect of animation synthesis. First, our system lets users select an object with the same class from the first video frame. We then employ random forests as learning algorithm to retrieve from a video the object which users want to insert into an artwork. Second, we utilize style transferring, which enables the video frames to be consistent with the style of painting. At last, we use the seamless image cloning algorithm to yield seamless synthesizing result. Our approach allows different users to synthesize animating paintings up to their own preferences. The resulting work not only maintains the original author's painting style, but also generates a variety of artistic conception for people to enjoy.

Dissertação de Mestrado em Design e Multimédia apresentada à Faculdade de Ciências e Tecnologia
Apesar das comuns aplicações para navegação, os mapas geográficos possuem uma inegável qualidade estética independente de fins utilitários que é, por isso, capaz de os apropriar a contextos estéticos, decorativos e focados no design. Os dados e ferramentas emergentes da constante evolução tecnológica popularizam e facilitam a utilização de mapas focados na estética e, desse modo, surgem novas oportunidades para exploração visual no contexto de mapas geográficos. Técnicas de design computacional e generativo facilitam e promovem a criação de artefactos em maiores quantidades e mais customizáveis, experimentais, e diversos, exacerbando assim a já mencionada oportunidade para exploração visual de mapas geográficos. Esta dissertação foca-se na utilização de técnicas de design computacional como ferramentas para estilização e exploração visual de mapas geográficos focados na estética. Nesse contexto, o projeto toma como objetivo o desenvolvimento de um sistema web capaz de gerar mapas estilizados de qualquer região geográfica do mundo. O sistema utiliza dados geográficos de acesso público para desenhar mapas de qualquer região e neural style transfer para os estilizar, permitindo assim a qualquer utilizador gerar mapas estilizados de qualquer lugar, baseados no estilo visual de qualquer imagem.Em primeiro lugar, é introduzida uma framework computacional que engloba os componentes e funcionalidades que permitem gerar mapas estilizados a partir de um conjunto de parâmetros. Por detrás desta framework está uma abordagem baseada em tiles que foi conceptualizada e implementada de forma a permitir adaptar qualquer implementação de neural style transfer para fins de estilização de mapas. De seguida, é apresentado o sistema web público que combina a framework de estilização com uma interface gráfica intuitiva de forma a permitir qualquer pessoa aceder e explorar o método de estilização de mapas aqui desenvolvido. A aplicabilidade dos mapas gerados é explorada e ilustrada, levando-nos a concluir que estes são uma forma viável de criar conexões a áreas geográficas de forma esteticamente apelativa, algo que pode acontecer através da sua aplicação em contextos estritamente decorativos ou mais significativos e focados no design.
Despite regular navigational applications, geographic maps possess an undeniable aesthetic quality independent from utilitarian purposes and thus able to propel them into decorative, aesthetic and design-focused contexts. The data and tools provided by the advancement of technology popularise and expedite the usage of aesthetic-focused maps and thereby open new opportunities for the visual exploration of geographic maps. Generative and computational design techniques aid and promote the creation of more numerous, customisable, diverse, and experimental outputs, thus exacerbating the aforementioned opportunity for visual exploration in the context of geographic maps.This dissertation focuses on the application of computational design techniques as a tool for map stylisation and the exploration of aesthetic-focused geographic maps. Within that context, the project centres on the development of a web-based system able to generate stylised maps of any region in the world. The system uses open-access geographic data to render maps of any selected area and a neural style transfer technique to stylise them, thus allowing users to generate stylised maps of anywhere, based on the visual style of any image file. We first introduce a stand-alone computational framework that comprises the components and functionalities tasked with generating stylised maps based on a set of parameters. At the core of this framework is a tile-based approach we conceptualised and implemented to successfully adapt any implementation of neural style transfer for the purposes of map stylisation. Then, we present a public web-based system that combines the stand-alone framework with an intuitive user interface, thereby enabling anyone to access and explore our approach to map stylisation. Finally, the real-world applicability of the generated stylised maps is explored and illustrated, leading us to conclude that they are a viable way to aesthetically establish connections to geographic places, which can happen through their application in purely decorative or more meaningful design contexts.