Thèses sur le sujet « No-reference metrics »

Images are playing a more and more important role in sharing, expressing, mining and exchanging information in our daily lives. Now we can all easily capture and share images anywhere and anytime. Since digital images are subject to a wide variety of distortions during acquisition, processing, compression, storage, transmission and reproduction; it becomes necessary to assess the Image Quality. In this thesis, starting from an organized overview of available Image Quality Assessment methods, some original contributions in the framework of No-reference image quality metrics are described.

Tese (doutorado)—Universidade de Brasília, Instituto de Ciências Exatas, Departamento de Ciência da Computação, 2017.
Submitted by Raquel Almeida (raquel.df13@gmail.com) on 2017-06-22T19:03:58Z No. of bitstreams: 1 2017_AlexandreFienodaSilva.pdf: 5179649 bytes, checksum: de1d53930e22f809bd34322d5c5270d0 (MD5)
Approved for entry into archive by Raquel Viana (raquelviana@bce.unb.br) on 2017-10-05T17:04:26Z (GMT) No. of bitstreams: 1 2017_AlexandreFienodaSilva.pdf: 5179649 bytes, checksum: de1d53930e22f809bd34322d5c5270d0 (MD5)
Made available in DSpace on 2017-10-05T17:04:26Z (GMT). No. of bitstreams: 1 2017_AlexandreFienodaSilva.pdf: 5179649 bytes, checksum: de1d53930e22f809bd34322d5c5270d0 (MD5) Previous issue date: 2017-10-05
Um dos principais fatores para a redução da qualidade do conteúdo visual, em sistemas de imagem digital, são a presença de degradações introduzidas durante as etapas de processamento de sinais. Contudo, medir a qualidade de um vídeo implica em comparar direta ou indiretamente um vídeo de teste com o seu vídeo de referência. Na maioria das aplicações, os seres humanos são o meio mais confiável de estimar a qualidade de um vídeo. Embora mais confiáveis, estes métodos consomem tempo e são difíceis de incorporar em um serviço de controle de qualidade automatizado. Como alternativa, as métricas objectivas, ou seja, algoritmos, são geralmente usadas para estimar a qualidade de um vídeo automaticamente. Para desenvolver uma métrica objetiva é importante entender como as características perceptuais de um conjunto de artefatos estão relacionadas com suas forças físicas e com o incômodo percebido. Então, nós estudamos as características de diferentes tipos de artefatos comumente encontrados em vídeos comprimidos (ou seja, blocado, borrado e perda-de-pacotes) por meio de experimentos psicofísicos para medir independentemente a força e o incômodo desses artefatos, quando sozinhos ou combinados no vídeo. Nós analisamos os dados obtidos desses experimentos e propomos vários modelos de qualidade baseados nas combinações das forças perceptuais de artefatos individuais e suas interações. Inspirados pelos resultados experimentos, nós propomos uma métrica sem-referência baseada em características extraídas dos vídeos (por exemplo, informações DCT, a média da diferença absoluta entre blocos de uma imagem, variação da intensidade entre pixels vizinhos e atenção visual). Um modelo de regressão não-linear baseado em vetores de suporte (Support Vector Regression) é usado para combinar todas as características e estimar a qualidade do vídeo. Nossa métrica teve um desempenho muito melhor que as métricas de artefatos testadas e para algumas métricas com-referência (full-reference).
The main causes for the reducing of visual quality in digital imaging systems are the unwanted presence of degradations introduced during processing and transmission steps. However, measuring the quality of a video implies in a direct or indirect comparison between test video and reference video. In most applications, psycho-physical experiments with human subjects are the most reliable means of determining the quality of a video. Although more reliable, these methods are time consuming and difficult to incorporate into an automated quality control service. As an alternative, objective metrics, i.e. algorithms, are generally used to estimate video quality quality automatically. To develop an objective metric, it is important understand how the perceptual characteristics of a set of artifacts are related to their physical strengths and to the perceived annoyance. Then, to study the characteristics of different types of artifacts commonly found in compressed videos (i.e. blockiness, blurriness, and packet-loss) we performed six psychophysical experiments to independently measure the strength and overall annoyance of these artifact signals when presented alone or in combination. We analyzed the data from these experiments and proposed several models for the overall annoyance based on combinations of the perceptual strengths of the individual artifact signals and their interactions. Inspired by experimental results, we proposed a no-reference video quality metric based in several features extracted from the videos (e.g. DCT information, cross-correlation of sub-sampled images, average absolute differences between block image pixels, intensity variation between neighbouring pixels, and visual attention). A non-linear regression model using a support vector (SVR) technique is used to combine all features to obtain an overall quality estimate. Our metric performed better than the tested artifact metrics and for some full-reference metrics.

Dissertação (mestrado)—Universidade de Brasília, Faculdade de Tecnologia, Departamento de Engenharia Elétrica, 2017.
Submitted by Raquel Almeida (raquel.df13@gmail.com) on 2017-05-26T21:27:49Z No. of bitstreams: 1 2017_DarioDanielRibeiroMorais.pdf: 11590245 bytes, checksum: 2daecd1489fc94cb4c8afb7736e1181f (MD5)
Approved for entry into archive by Raquel Viana (raquelviana@bce.unb.br) on 2017-05-29T23:07:11Z (GMT) No. of bitstreams: 1 2017_DarioDanielRibeiroMorais.pdf: 11590245 bytes, checksum: 2daecd1489fc94cb4c8afb7736e1181f (MD5)
Made available in DSpace on 2017-05-29T23:07:11Z (GMT). No. of bitstreams: 1 2017_DarioDanielRibeiroMorais.pdf: 11590245 bytes, checksum: 2daecd1489fc94cb4c8afb7736e1181f (MD5) Previous issue date: 2017-05-29
Este trabalho visa desenvolver uma métrica híbrida de qualidade de vídeo sem referência para aplicações de transmissão digital, que leva em consideração três tipos de artefatos: perda de pacotes, blocado e borrado. As características desses artefatos são extraídas a partir das sequências de vídeo a fim de quantificar a força desses artefatos. A avaliação de perda de pacotes é dividida em 2 etapas: detecção e medição. As avaliações de blocado e borrado seguem referências da literatura. Depois de obter as características dos três tipos de artefatos, um processo de aprendizado de máquina (SVR) é utilizado para estimar a nota de qualidade prevista a partir das características extraídas. Os resultados obtidos com a métrica proposta foram comparados com os resultados obtidos com outras três métricas disponíveis na literatura (duas métricas NR de perda de pacotes e 1 métrica FR) e eles são promissores. A métrica proposta é cega, rápida e confiável para ser usada em cenários em tempo real.
This work aims to develop a hybrid no-reference video quality metric for digital transmission applications, which takes into account three types of artifacts: packet-loss, blockiness and bluriness. Features are extracted from the video sequences in order to quantity the strength of these three artifacts. The assessment of the packet-loss strength is performed in 2 stages: detection and measurement. The assessment of the strength of blockiness and blussiness follow references from literature. After obtaining the features from these three types of artifacts, a machine learning algorithm ( the support vector regression technique), is used to estimate the predicted quality score from the extracted features. The results obtained with the proposed metric were compared with the results obtained with three other metrics available in the literature (two NR packet-loss metrics and one FR metric). The proposed metric is blind, fast, and reliable to be used in real-time scenarios.

Les systèmes d'identification de personnes basés sur le visage deviennent de plus en plus répandus et trouvent des applications très variées, en particulier dans le domaine de la vidéosurveillance. Or, dans ce contexte, les performances des algorithmes de reconnaissance faciale dépendent largement des conditions d'acquisition des images, en particulier lorsque la pose varie mais également parce que les méthodes d'acquisition elles mêmes peuvent introduire des artéfacts. On parle principalement ici de maladresse de mise au point pouvant entraîner du flou sur l'image ou bien d'erreurs liées à la compression et faisant apparaître des effets de blocs. Le travail réalisé au cours de la thèse porte donc sur la reconnaissance de visages à partir d'images acquises à l'aide de caméras de vidéosurveillance, présentant des artéfacts de flou ou de bloc ou bien des visages avec des poses variables. Nous proposons dans un premier temps une nouvelle approche permettant d'améliorer de façon significative la reconnaissance des visages avec un niveau de flou élevé ou présentant de forts effets de bloc. La méthode, à l'aide de métriques spécifiques, permet d'évaluer la qualité de l'image d'entrée et d'adapter en conséquence la base d'apprentissage des algorithmes de reconnaissance. Dans un second temps, nous nous sommes focalisés sur l'estimation de la pose du visage. En effet, il est généralement très difficile de reconnaître un visage lorsque celui-ci n'est pas de face et la plupart des algorithmes d'identification de visages considérés comme peu sensibles à ce paramètre nécessitent de connaître la pose pour atteindre un taux de reconnaissance intéressant en un temps relativement court. Nous avons donc développé une méthode d'estimation de la pose en nous basant sur des méthodes de reconnaissance récentes afin d'obtenir une estimation rapide et suffisante de ce paramètre
The person identification systems based on face recognition are becoming increasingly widespread and are being used in very diverse applications, particularly in the field of video surveillance. In this context, the performance of the facial recognition algorithms largely depends on the image acquisition context, especially because the pose can vary, but also because the acquisition methods themselves can introduce artifacts. The main issues are focus imprecision, which can lead to blurred images, or the errors related to compression, which can introduce the block artifact. The work done during the thesis focuses on facial recognition in images taken by video surveillance cameras, in cases where the images contain blur or block artifacts or show various poses. First, we are proposing a new approach that allows to significantly improve facial recognition in images with high blur levels or with strong block artifacts. The method, which makes use of specific noreference metrics, starts with the evaluation of the quality level of the input image and then adapts the training database of the recognition algorithms accordingly. Second, we have focused on the facial pose estimation. Normally, it is very difficult to recognize a face in an image taken from another viewpoint than the frontal one and the majority of facial identification algorithms which are robust to pose variation need to know the pose in order to achieve a satisfying recognition rate in a relatively short time. We have therefore developed a fast and satisfying pose estimation method based on recent recognition techniques

Submitted by Daniele Amaral (daniee_ni@hotmail.com) on 2016-09-21T19:25:25Z No. of bitstreams: 1 DissAOL.pdf: 6820225 bytes, checksum: 00793d2a933ccdc18ce28c12bf53f7ac (MD5)
Approved for entry into archive by Ronildo Prado (ronisp@ufscar.br) on 2016-09-28T19:38:02Z (GMT) No. of bitstreams: 1 DissAOL.pdf: 6820225 bytes, checksum: 00793d2a933ccdc18ce28c12bf53f7ac (MD5)
Approved for entry into archive by Ronildo Prado (ronisp@ufscar.br) on 2016-09-28T19:38:13Z (GMT) No. of bitstreams: 1 DissAOL.pdf: 6820225 bytes, checksum: 00793d2a933ccdc18ce28c12bf53f7ac (MD5)
Made available in DSpace on 2016-09-28T19:43:11Z (GMT). No. of bitstreams: 1 DissAOL.pdf: 6820225 bytes, checksum: 00793d2a933ccdc18ce28c12bf53f7ac (MD5) Previous issue date: 2015-10-05
Não recebi financiamento
This research aims to propose activities for teachers using the Geogebra software, especially for teachers from the SESI-SP School Network in order to assist them in the teaching methodology, with teachers' work plan and, in addition, aiming to more significant and dynamic classes, in order to allow students reach their teaching and learning expectations, formulate valid arguments, make conjectures and justify their reasoning. The activities were applied by teachers of SESI-SP School Network to the students of 9th grade of elementary school, in anticipation of teaching and learning through “Similarity”, addressing Theorem of Thales, Metrics Relations in the Rectangle Triangle and Pythagoras Theorem. The results were analyzed and discussed, reporting the challenges and conclusions raised by the students during the activities while working with the Geogebra software and also based on the feedback provided by the teachers and the opinion of the analysts from SESI-SP School Network.
Esta pesquisa tem como objetivo principal propor atividades para os professores utilizando o software Geogebra, principalmente para os docentes da rede SESI-SP de Ensino, a fim de auxiliá-los na metodologia de ensino, no plano de trabalho, visando uma aula mais significativa e dinâmica, para que seus alunos atinjam as expectativas de ensino e aprendizagem, formulem argumentos válidos, façam conjecturas e justifiquem seus raciocínios. As atividades foram aplicadas por professores da rede SESI-SP de Ensino aos alunos do 9º ano do Ensino Fundamental, turma de 2014, na expectativa de ensino e aprendizagem de “Semelhança”, abordando Teorema de Tales, Relações Métricas no Triângulo Retângulo e Teorema de Pitágoras. Os resultados foram analisados e discutidos, relatando as dificuldades e conclusões apresentadas pelos alunos em desenvolver as atividades trabalhando com o software Geogebra, baseado nas devolutivas dos professores envolvidos e o parecer feito pelos analistas educacionais da Rede SESI-SP de Ensino.

This thesis presents a way of performing objective video quality analyses in order to point out faults in the hardware of a video system that uses analogue video transmission technologies. The approach focuses on performing simple digital processing and analyses of the video data coherently using an FPGA. Several metrics that correlates with specific distortions are developed. These metrics give good indications of the state of the video system components. The algorithms are tested using MATLAB and mapped to an FPGA. The key components are implemented and verified in VHDL, and synthesized for an Altera Cyclone II FPGA. The thesis concludes that the proposed system has the ability to discover board-level faults in a video system that utilizes an FPGA and analogue video transmission. The system also has the ability to supplement external quality assessment systems in most cases, and function as a good alternative in cases where a quick and simple assessment of a video system is desired.

This diploma thesis deals with objective video quality assessments without reference. Some of the basics of video quality evaluation are described. Also some basic conditions for objective video quality metrics are introduced. The main focus of the thesis are no-reference approaches. Thesis tries to describe basic methods for seeking distortion in video. The difference between spatial-domain oriented and spectral-domain oriented metrics is analyzed. We also describe design of tool for measuring objective video quality in Matlab environment. We then designed and implemented a bit-stream oriented metric for estimation of the PSNR of H.264 coded sequences. Finally, we created a database of video sequences and we held objective tests. Results were compared with results from subjective measurements.

A causa della limitata estensione del campo di vista del sistema, con un microscopio ottico tradizionale non è possibile acquisire una singola immagine completamente a fuoco se l’oggetto è caratterizzato da una profondità non trascurabile. Fin dagli anni ’70, il problema dell’estensione della profondità di fuoco è stato ampiamente trattato in letteratura e molti metodi sono stati proposti per superare questo limite. Tuttavia, è molto difficile riuscire a decretare quale metodo risulti essere il migliore in una specifica applicazione, a causa della mancanza di una metrica validata e adatta ad essere utilizzata anche in casi reali, dove generalmente non si ha a disposizione un’immagine di riferimento da considerare come “la verità” (ground truth). L’Universal Quality Index (UQI) è ampiamente utilizzato in letteratura per valutare la qualità dei risultati in processi di elaborazione di immagini. Tuttavia, per poter calcolare questo indice è necessaria una ground truth. In effetti, sono state proposte in letteratura alcune estensioni dell’UQI per valutare il risultato dei metodi di fusione anche senza immagine di riferimento, ma nessuna analisi esaustiva è stata proposta per dimostrare la loro equivalenza con l’UQI standard nel valutare la qualità di un’immagine. In questo lavoro di Tesi, partendo dai limiti dei metodi attualmente utilizzati per l'estensione della profondità di campo di un microscopio, ed esposti in letteratura, per prima cosa è stato proposto un nuovo metodo, basato su approccio spaziale e fondato su analisi locale del segnale appositamente filtrato. Attraverso l’uso di sequenze di immagini sintetiche, delle quali si conosce la ground truth, è stato dimostrato, utilizzando metriche comuni in image processing, che il metodo proposto è in grado di superare le performance dei metodi allo stato dell'arte. In seguito, attraverso una serie di esperimenti dedicati, è stato provato che metriche proposte e ampiamente utilizzate in letteratura come estensione dell'UQI per valutare la qualità delle immagini prodotte da processi di fusione, sebbene dichiarate essere sue estensioni, non sono in grado di effettuare una valutazione quale quella che farebbe l'UQI standard. E’ quindi stato proposto e validato un nuovo approccio di valutazione che si è dimostrato in grado di classificare i metodi di fusione testati così come farebbe l’UQI standard, ma senza richiedere un’immagine di riferimento. Infine, utilizzando sequenze di immagini reali acquisite a differenti profondità di fuoco e l’approccio di valutazione validato, è stato dimostrato che il metodo proposto per l’estensione della profondità di campo risulta sempre migliore, o almeno equivalente, ai metodi allo stato dell’arte.

碩士
中原大學
電子工程研究所
101
In this paper, we propose no-reference compressed video quality models to predict the full-reference PSNR and SSIM metrics for AVC (Advanced Video Coding, H.264) and HEVC (High Efficiency Video Coding) encoded videos. The model we used is support vector regression (SVR) that transforms the features into higher-dimensional space to achieve better prediction performance. We use only encoding decisions made during motion estimation to perform the prediction, and do not need the information from pixel domain, so it is very efficient. We show that the SVR model containing the factors related to the statistics about block partitions in a frame can predict the video quality well for AVC videos (0.81 correlation for PSNR, and 0.79 for SSIM). This approach provides even higher prediction performance for HEVC videos (0.91 for PSNR and 0.89 for SSIM) due to its more complex partition decisions than AVC; the improvement of HEVC model over AVC model is 12% for PSNR prediction, and 11% for SSIM prediction. Other model performance measurement RMSE and R2 are also provided to support the results. This paper demonstrates that, other than increasing the encoding efficiency, the complex encoding parameters in HEVC can provide more information about original frames compared to AVC.

碩士
國立臺灣海洋大學
資訊工程學系
104
To objectively quantify the perceived blurriness of an image is useful and impertant in various image processing applications, in particular image quality assessment without the reference image is more challenging. In this paper, we present a blind image blur metric adapted to different image content. The approach is originated from the wavelet characterization of smoothness of Besov function spaces. We then imitate the human-perceived blurriness through Besov smoothness model. The metric utilizes the descending rate of Besov norm of the re-blurred images to estimate the blurred image quality. The performance of the proposed framework is evaluated on four public image quality databases. The experimental results show that our method can produce blur index highly consistent with subjective evaluations. Keywords: Blind quality assessment, blur image metric, wavelet transform, Besov space.

碩士
逢甲大學
通訊工程所
99
Image quality is good or bad, depending on the natural, contrast, color degree, different factors such as ambiguity, all the factors to be applied to determine the image quality is very difficult, so the main goal of this study is to estimate the blur of the previous laboratory the proposed method estimates uniform in the direction blurred images and DMOS are more highly correlated, but non-isotropic in the estimation of uniform image blur type, relevance to much lower rates, in order to improve the deficiency, we the original filter down into several directions, in all directions out of the minimum estimated value, will be able to estimate the non-uniform isotropic blurred image. In addition to this vague image in the regional architecture, ways to cut blocks, the blocks can be obtained fuzzy strong.

碩士
國立中興大學
電機工程學系所
105
Perceptual quality plays an irreplaceable importance in viewing stereoscopic 3d images. Generally, worse quality stereoscopic 3D images will cause the viewer’s eyes tolerate fatigue, painfulness or feel dizziness, headache. Therefore, in this study, we propose a no-reference metric for stereoscopic image quality of experience (QoE) to evaluate the visual discomfort when the viewers view stereoscopic images. We develop two regression models in machine learning (ML), support vector machine (SVM) and random forest (RF), to assess the scores of visual discomfort and then compare the performance between two models. We test our method on the publicly available EPFL 3D image database and IEEE-SA stereoscopic image databases. First, the disparity of stereoscopic pairs is calculated and the depth information, called the depth-disparity map, is obtained from the resulting disparity map through Otsu’s algorithm. Next, four kinds of features are extracted based on the pixel values and distribution of depth-disparity map to build the input data and then use above-mentioned two regression models to analyze the data. Finally, the correlation between the predicted scores obtained from the proposed metric and the subjective scores provided by the databases is calculated. The experimental results show that the proposed metric achieves an impressive performance comparing with current state-of-the-art methods.

碩士
國立中正大學
資訊工程研究所
106
Nowadays, Internet is booming and the perception of video quality by video providers and users is becoming more important, but limit by the bandwidth of network transmission. No reference video quality computation is the best and well-known in three types of video quality assessment metrics. In this study, the proposed video quality computation metric is based on no reference and extracted spatial, temporal, transform, and spatiotemporal features as the basis for predicting quality scores. First, edge detection and blockiness are extracted as the spatial features and difference of luminance and motion are extracted as temporal features. The pairwise products of discrete cosine transform and wavelet transform are extracted to enhance the center point pixel and surrounding neighbor pixels, and are regarded as transform features. Considering that spatial and temporal information can extracted simultaneously, the statistical properties of trajectory and three-dimensional discrete cosine transform are taken as spatiotemporal features. Finally, support vector regression is utilized to predict the final quality score. This experiment using LIVE video quality assessment database and experimental results show that the results have better results than other existing metrics.

碩士
國立中正大學
資訊工程研究所
106
In recent year, 3D technology application provides a new viewing experience and has become more and more widespread. Due to the reason mentioned above, humans will pay more attention on stereoscopic video quality. In other words, it is necessary to develop the stereoscopic video quality assessment approaches. Full-reference and reduced-reference stereoscopic video quality assessment methods usually obtain better performance since these approaches can make use of the information of original videos. However, it is hard to get original videos when transmitting. Hence, no-reference stereoscopic video quality assessment technology is mainly focused in this study. First, features from four domains, including spatial, depth, transform, and spatiotemporal features are extracted. On the spatial domain, blockiness, cyclopean view, binocular rivalry, cross entropy, and edge are extracted. On the depth domain, disparity saliency, depth structure, NSS, and depth entropy are extracted. On the transform domain, discrete wavelet transform (DWT) and contourlet transform information are extracted. On the spatiotemporal domain, depth motion and 3D-DCT information are extracted. The feature vectors from the left-view and right-view videos are averaged and represented as statistical feature and normalize to the same distribution. Then, support vector regression (SVR) is applied to measure the stereoscopic video quality score. Finally, experimental results show that the proposed approach is better than the other NR approach on NAMA3DS1_COSPAD1 database.

碩士
大同大學
通訊工程研究所
98
Automatic picture quality control is a very important function to the implementation of many digital information systems, such as the broadcast of TV station, medical imaging systems, image/video transmission of networks. Automatic picture quality control system depends on the effectiveness of the objective quality assessment metric. Since original images can’t be conveniently acquired, an ideal objective quality assessment metric must be enabled to evaluate image quality without examining the original image as a reference and give the result highly consistent with subjective visual perception. The purpose of this thesis is expect to present a novel no-reference image quality metric which can be highly consistent with the perception of human visual system. Based on the idea that the quality of image is related to the variation of image features which generated from blurring/sharpening image to propose an objective quality assessment metric. In this research, image features of spatial domain and frequency domain are used to assess the quality of test images. Because luminance signal is the most important constituent of color image, so the luminance signal values of test images are calculated, and then the luminance signal values are transformed to the coefficients of frequency domain by Haar transform. After blurred/ sharpened images are compared with test images, the variations of signal energy and signal energy center which in frequency domain are calculated. In addition, the variations of entropy of spatial domain are calculated. Objective metric scores are based on above variation values to evaluate. Finally, for improving this method is effective, objective metric scores are compared with subjective metric scores that from experiment. The proposed no-reference image quality metric of this thesis is expected to give results which correlate to subjective visual perception, not affected by external effects and doesn't need original image.

碩士
逢甲大學
通訊工程所
98
The quality of an image depends on various attributes such as blurriness, naturalness, colorfulness, and contrast, etc. To develop an image quality metric by incorporating all attributes of images, meanwhile, without referring to the original images, is a difficult task. Hence, the purpose of this thesis is to develop a no-reference objective image blur metric. Two blurriness measures are proposed: one is edge-based and the other is energy-based. Both blurriness measuring methods compute the metrics based on the output of a filter bank for a given image. A snug frame composed of the Gaussian derivative wavelets is applied to construct the filter bank. The output of the filter bank not only contains the complete information of an input image but also manifests prominent image features. Experimental results show that both metrics predict well in isotropic and uniform blur case. The performances of the metrics are comparable to that of human subjects.