Academic literature on the topic 'Domain adaptation, domain-shift, image classification, neural networks'

When large-scale annotated data are not available for certain image classification tasks, training a deep convolutional neural network model becomes challenging. Some recent domain adaptation methods try to solve this problem using generative adversarial networks and have achieved promising results. However, these methods are based on a shared latent space assumption and they do not consider the situation when shared high level representations in different domains do not exist or are not ideal as they assumed. To overcome this limitation, we propose a neural network structure called coupled generative adversarial autoencoders (CGAA) that allows a pair of generators to learn the high-level differences between two domains by sharing only part of the high-level layers. Additionally, by introducing a class consistent loss calculated by a stand-alone classifier into the generator optimization, our model is able to generate class invariant style-transferred images suitable for classification tasks in domain adaptation. We apply CGAA to several domain transferred image classification scenarios including several benchmark datasets. Experiment results have shown that our method can achieve state-of-the-art classification results.

The use of Convolutional Neural Networks (CNNs) to solve Domain Adaptation (DA) image classification problems in the context of remote sensing has proven to provide good results but at high computational cost. To avoid this problem, a deep learning network for DA in remote sensing hyperspectral images called TCANet is proposed. As a standard CNN, TCANet consists of several stages built based on convolutional filters that operate on patches of the hyperspectral image. Unlike the former, the coefficients of the filter are obtained through Transfer Component Analysis (TCA). This approach has two advantages: firstly, TCANet does not require training based on backpropagation, since TCA is itself a learning method that obtains the filter coefficients directly from the input data. Second, DA is performed on the fly since TCA, in addition to performing dimensional reduction, obtains components that minimize the difference in distributions of data in the different domains corresponding to the source and target images. To build an operating scheme, TCANet includes an initial stage that exploits the spatial information by providing patches around each sample as input data to the network. An output stage performing feature extraction that introduces sufficient invariance and robustness in the final features is also included. Since TCA is sensitive to normalization, to reduce the difference between source and target domains, a previous unsupervised domain shift minimization algorithm consisting of applying conditional correlation alignment (CCA) is conditionally applied. The results of a classification scheme based on CCA and TCANet show that the DA technique proposed outperforms other more complex DA techniques.

Deep neural networks can learn powerful representations from massive amounts of labeled data; however, their performance is unsatisfactory in the case of large samples and small labels. Transfer learning can bridge between a source domain with rich sample data and a target domain with only a few or zero labeled samples and, thus, complete the transfer of knowledge by aligning the distribution between domains through methods, such as domain adaptation. Previous domain adaptation methods mostly align the features in the feature space of all categories on a global scale. Recently, the method of locally aligning the sub-categories by introducing label information achieved better results. Based on this, we present a deep fuzzy domain adaptation (DFDA) that assigns different weights to samples of the same category in the source and target domains, which enhances the domain adaptive capabilities. Our experiments demonstrate that DFDA can achieve remarkable results on standard domain adaptation datasets.

Abstract Background and objective. Optical coherence tomography (OCT) is one of the most used retinal imaging modalities in the clinic as it can provide high-resolution anatomical images. The huge number of OCT images has significantly advanced the development of deep learning methods for automatic lesion detection to ease the doctor’s workload. However, it has been frequently revealed that the deep neural network model has difficulty handling the domain discrepancies, which widely exist in medical images captured from different devices. Many works have been proposed to solve the domain shift issue in deep learning tasks such as disease classification and lesion segmentation, but few works focused on lesion detection, especially for OCT images. Methods. In this work, we proposed a faster-RCNN based, unsupervised domain adaptation model to address the lesion detection task in cross-device retinal OCT images. The domain shift is minimized by reducing the image-level shift and instance-level shift at the same time. We combined a domain classifier with a Wasserstein distance critic to align the shifts at each level. Results. The model was tested on two sets of OCT image data captured from different devices, obtained an average accuracy improvement of more than 8% over the method without domain adaptation, and outperformed other comparable domain adaptation methods. Conclusion. The results demonstrate the proposed model is more effective in reducing the domain shift than advanced methods.

Deep neural networks excel at learning from large-scale labeled training data, but cannot well generalize the learned knowledge to new domains or datasets. Domain adaptation studies how to transfer models trained on one labeled source domain to another sparsely labeled or unlabeled target domain. In this paper, we investigate the unsupervised domain adaptation (UDA) problem in image emotion classification. Specifically, we develop a novel cycle-consistent adversarial model, termed CycleEmotionGAN, by enforcing emotional semantic consistency while adapting images cycleconsistently. By alternately optimizing the CycleGAN loss, the emotional semantic consistency loss, and the target classification loss, CycleEmotionGAN can adapt source domain images to have similar distributions to the target domain without using aligned image pairs. Simultaneously, the annotation information of the source images is preserved. Extensive experiments are conducted on the ArtPhoto and FI datasets, and the results demonstrate that CycleEmotionGAN significantly outperforms the state-of-the-art UDA approaches.

Unsupervised domain adaptation involves knowledge transfer from a labeled source to unlabeled target domains to assist target learning tasks. A critical aspect of unsupervised domain adaptation is the learning of more transferable and distinct feature representations from different domains. Although previous investigations, using, for example, CNN-based and auto-encoder-based methods, have produced remarkable results in domain adaptation, there are still two main problems that occur with these methods. The first is a training problem for deep neural networks; some optimization methods are ineffective when applied to unsupervised deep networks for domain adaptation tasks. The second problem that arises is that redundancy of image data results in performance degradation in feature learning for domain adaptation. To address these problems, in this paper, we propose an unsupervised domain adaptation method with a stacked convolutional sparse autoencoder, which is based on performing layer projection from the original data to obtain higher-level representations for unsupervised domain adaptation. More specifically, in a convolutional neural network, lower layers generate more discriminative features whose kernels are learned via a sparse autoencoder. A reconstruction independent component analysis optimization algorithm was introduced to perform individual component analysis on the input data. Experiments undertaken demonstrated superior classification performance of up to 89.3% in terms of accuracy compared to several state-of-the-art domain adaptation methods, such as SSRLDA and TLMRA.

Crop losses due to plant diseases isa serious problem for the farming sector of agricultureand the economy. Therefore, a multi-functional Plant Disease Detection Platform (PDDP) was developed in the LIT JINR. Deep learning techniques are successfully used in PDDP to solve the problem of recognizing plant diseases from photographs of their leaves. However, such methods require a large training dataset. At the same time, there are number of methods used to solve classification problems in cases of a small training dataset, asfor example,domain adaptation(DA)methods.In this paper, a comparative study of three DA methods is performed:Domain-Adversarial Training of Neural Networks (DANN), two-steps transfer learning and Unsupervised Domain Adaptation with Deep Metric Learning (M-ADDA).The advantage of the M-ADDA methodwas shown, which allowed toachieve 92% ofclassification accuracy.

The plastic modifications in synaptic connectivity is primarily from changes triggered by neuromodulated dopamine signals. These activities are controlled by neuromodulation, which is itself under the control of the brain. The subjective brain’s self-modifying abilities play an essential role in learning and adaptation. The artificial neural networks with neuromodulated plasticity are used to implement transfer learning in the image classification domain. In particular, this has application in image detection, image segmentation, and transfer of learning parameters with significant results. This paper proposes a novel approach to enhance transfer learning accuracy in a heterogeneous source and target, using the neuromodulation of the Hebbian learning principle, called NDHTL (Neuromodulated Dopamine Hebbian Transfer Learning). Neuromodulation of plasticity offers a powerful new technique with applications in training neural networks implementing asymmetric backpropagation using Hebbian principles in transfer learning motivated CNNs (Convolutional neural networks). Biologically motivated concomitant learning, where connected brain cells activate positively, enhances the synaptic connection strength between the network neurons. Using the NDHTL algorithm, the percentage of change of the plasticity between the neurons of the CNN layer is directly managed by the dopamine signal’s value. The discriminative nature of transfer learning fits well with the technique. The learned model’s connection weights must adapt to unseen target datasets with the least cost and effort in transfer learning. Using distinctive learning principles such as dopamine Hebbian learning in transfer learning for asymmetric gradient weights update is a novel approach. The paper emphasizes the NDHTL algorithmic technique as synaptic plasticity controlled by dopamine signals in transfer learning to classify images using source-target datasets. The standard transfer learning using gradient backpropagation is a symmetric framework. Experimental results using CIFAR-10 and CIFAR-100 datasets show that the proposed NDHTL algorithm can enhance transfer learning efficiency compared to existing methods.

Learning has long been and will continue to be a key issue in intelligent algorithms and systems design. Emulating the behavior and mechanisms of human learning by machines at such high levels as symbolic processing and such low levels as neuronal processing has long been a dominant interest among researchers worldwide. Neural networks, fuzzy logic, and evolutionary algorithms represent the three most active research areas. With advanced theoretical studies and computer technology, many promising algorithms and systems using these techniques have been designed and implemented for a wide range of applications. This Special Issue presents seven papers on learning in intelligent algorithms and systems design from researchers in Japan, China, Australia, and the U.S. <B>Neural Networks:</B> Emulating low-level human intelligent processing, or neuronal processing, gave birth of artificial neural networks more than five decades ago. It was hoped that devices based on biological neural networks would possess characteristics of the human brain. Neural networks have reattracted researchers' attention since the late 1980s when back-propagation algorithms were used to train multilayer feed-forward neural networks. In the last decades, we have seen promising progress in this research field yield many new models, learning algorithms, and real-world applications, evidenced by the publication of new journals in this field. <B>Fuzzy Logic:</B> Since L. A. Zadeh introduced fuzzy set theory in 1965, fuzzy logic has increasingly become the focus of many researchers and engineers opening up new research and problem solving. Fuzzy set theory has been favorably applied to control system design. In the last few years, fuzzy model applications have bloomed in image processing and pattern recognition. <B>Evolutionary Algorithms:</B> Evolutionary optimization algorithms have been studied over three decades, emulating natural evolutionary search and selection so powerful in global optimization. The study of evolutionary algorithms includes evolutionary programming (EP), evolutionary strategies (ESs), genetic algorithms (GAs), and genetic programming (GP). In the last few years, we have also seen multiple computational algorithms combined to maximize system performance, such as neurofuzzy networks, fuzzy neural networks, fuzzy logic and genetic optimization, neural networks, and evolutionary algorithms. This Special Issue also includes papers that introduce combined techniques. <B>Wang</B> et al present an improved fuzzy algorithm for enhanced eyeground images. Examination of the eyeground image is effective in diagnosing glaucoma and diabetes. Conventional eyeground image quality is usually too poor for doctors to obtain useful information, so enhancement is required to eliminate this. Due to details and uncertainties in eyeground images, conventional enhancement such as histogram equalization, edge enhancement, and high-pass filters fail to achieve good results. Fuzzy enhancement enhances images in three steps: (1) transferring an image from the spatial domain to the fuzzy domain; (2) conducting enhancement in the fuzzy domain; and (3) returning the image from the fuzzy domain to the spatial domain. The paper detailing this proposes improved mapping and fast implementation. <B>Mohammadian</B> presents a method for designing self-learning hierarchical fuzzy logic control systems based on the integration of evolutionary algorithms and fuzzy logic. The purpose of such an approach is to provide an integrated knowledge base for intelligent control and collision avoidance in a multirobot system. Evolutionary algorithms are used as in adaptation for learning fuzzy knowledge bases of control systems and learning, mapping, and interaction between fuzzy knowledge bases of different fuzzy logic systems. Fuzzy integral has been found useful in data fusion. <B>Pham and Wagner</B> present an approach based on the fuzzy integral and GAs to combine likelihood values of cohort speakers. The fuzzy integral nonlinearly fuses similarity measures of an utterance assigned to cohort speakers. In their approach, Gas find optimal fuzzy densities required for fuzzy fusion. Experiments using commercial speech corpus T146 show their approach achieves more favorable performance than conventional normalization. Evolution reflects the behavior of a society. <B>Puppala and Sen</B> present a coevolutionary approach to generating behavioral strategies for cooperating agent groups. Agent behavior evolves via GAs, where one genetic algorithm population is evolved per individual in the cooperative group. Groups are evaluated by pairing strategies from each population and best strategy pairs are stored together in shared memory. The approach is evaluated using asymmetric room painting and results demonstrate the superiority of shared memory over random pairing in consistently generating optimal behavior patterns. Object representation and template optimization are two main factors affecting object recognition performance. <B>Lu</B> et al present an evolutionary algorithm for optimizing handwritten numeral templates represented by rational B-spline surfaces of character foreground-background-distance distribution maps. Initial templates are extracted from training a feed-forward neural network instead of using arbitrarily chosen patterns to reduce iterations required in evolutionary optimization. To further reduce computational complexity, a fast search is used in selection. Using 1,000 optimized numeral templates, the classifier achieves a classification rate of 96.4% while rejecting 90.7% of nonnumeral patterns when tested on NIST Special Database 3. Determining an appropriate number of clusters is difficult yet important. <B>Li</B> et al based their approach based on rival penalized competitive learning (RPCL), addressing problems of overlapped clusters and dependent components of input vectors by incorporating full covariance matrices into the original RPCL algorithm. The resulting learning algorithm progressively eliminates units whose clusters contain only a small amount of training data. The algorithm is applied to determine the number of clusters in a Gaussian mixture distribution and to optimize the architecture of elliptical function networks for speaker verification and for vowel classification. Another important issue on learning is <B>Kurihara and Sugawara's</B> adaptive reinforcement learning algorithm integrating exploitation- and exploration-oriented learning. This algorithm is more robust in dynamically changing, large-scale environments, providing better performance than either exploitation- learning or exploration-oriented learning, making it is well suited for autonomous systems. In closing we would like to thank the authors who have submitted papers to this Special Issue and express our appreciation to the referees for their excellent work in reading papers under a tight schedule.

<p><strong>Abstract.</strong> Domain adaptation (DA) can drastically decrease the amount of training data needed to obtain good classification models by leveraging available data from a source domain for the classification of a new (target) domains. In this paper, we address deep DA, i.e. DA with deep convolutional neural networks (CNN), a problem that has not been addressed frequently in remote sensing. We present a new method for semi-supervised DA for the task of pixel-based classification by a CNN. After proposing an encoder-decoder-based fully convolutional neural network (FCN), we adapt a method for adversarial discriminative DA to be applicable to the pixel-based classification of remotely sensed data based on this network. It tries to learn a feature representation that is domain invariant; domain-invariance is measured by a classifier’s incapability of predicting from which domain a sample was generated. We evaluate our FCN on the ISPRS labelling challenge, showing that it is close to the best-performing models. DA is evaluated on the basis of three domains. We compare different network configurations and perform the representation transfer at different layers of the network. We show that when using a proper layer for adaptation, our method achieves a positive transfer and thus an improved classification accuracy in the target domain for all evaluated combinations of source and target domains.</p>

In the recent years, remote sensing has faced a huge evolution. The constantly growing availability of remote sensing data has opened up new opportunities and laid the foundations for many new challenges. The continuous space missions and new constellations of satellites allow in fact more and more frequent acquisitions, at increasingly higher spatial resolutions, and at an almost total coverage of the globe. The availability of such an huge amount data has highlighted the need for automatic techniques capable of processing the data and exploiting all the available information. Meanwhile, the almost unlimited potential of machine learning has changed the world we live in. Artificial neural Networks have break trough everyday life, with applications that include computer vision, speech processing, autonomous driving but which are also the basis of commonly used tools such as online search engines. However, the vast majority of such models are of the supervised type and therefore their applicability rely on the availability of an enormous quantity of labeled data available to train the models themselves. Unfortunately, this is not the case with remote sensing, in which the enormous amounts of data are opposed to the almost total absence of ground truth. The purpose of this thesis is to find the way to exploit the most recent deep learning techniques, defining a common thread between two worlds, those of remote sensing and deep learning, which is often missing. In particular, this thesis proposes three novel contributions which face current issues in remote sensing. The first one is related to multisensor image registration and combines generative adversarial networks and non-linear optimization of crosscorrelation-like functionals to deal with the complexity of the setting. The proposed method was proved able to outperform state of the art approaches. The second novel contribution faces one of the main issues in deep learning for remote sensing: the scarcity of ground truth data for semantic segmentation. The proposed solution combines convolutional neural networks and probabilistic graphical models, two very active areas in machine learning for remote sensing, and approximate a fully connected conditional random field. The proposed method is capable of filling part of the gap which separate a densely trained model from a weakly trained one. Then, the third approach is aimed at the classification of high resolution satellite images for climate change purposes. It consist of a specific formulation of an energy minimization which allows to fuse multisensor information and the application a markov random field in a fast and efficient way for global scale applications. The results obtained in this thesis shows how deep learning methods based on artificial neural networks can be combined with statistical analysis to overcome their limitations, going beyond the classic benchmark environments and addressing practical, real and large-scale application cases.

The availability of annotated image datasets and recent advances in supervised deep learning methods are enabling the end-to-end derivation of representative image features that can impact a variety of image analysis problems. These supervised methods use prior knowledge derived from labelled training data and approaches, for example, convolutional neural networks (CNNs) have produced impressive results in natural (photographic) image classification. CNNs learn image features in a hierarchical fashion. Each deeper layer of the network learns a representation of the image data that is higher level and semantically more meaningful. However, the accuracy and robustness of image features with supervised CNNs are dependent on the availability of large-scale labelled training data. In medical imaging, these large labelled datasets are scarce mainly due to the complexity of manual annotation and inter- and intra-observer variability in label assignment. The concept of ‘transfer learning’ – the adoption of image features from different domains, e.g., image features learned from natural photographic images – was introduced to address the lack of large amounts of labelled medical image data. These image features, however, are often generic and do not perform well in specific medical image analysis problems. An alternative approach was to optimise these features by retraining the generic features using a relatively small set of labelled medical images. This ‘fine-tuning’ approach, however, is not able to match the overall accuracy of learning image features directly from large collections of data that are specifically related to the problem at hand. An alternative approach is to use unsupervised feature learning algorithms to build features from unlabelled data, which then allows unannotated image archives to be used. Many unsupervised feature learning algorithms such as sparse coding (SC), auto-encoder (AE) and Restricted Boltzmann Machines (RBMs), however, have often been limited to learning low-level features such as lines and edges. In an attempt to address these limitations, in this thesis, we present several new unsupervised deep learning methods to learn semantic high-level features from unlabelled medical images to address the challenge of learning representative visual features in medical image analysis. We present two methods to derive non-linear and non-parametric models, which are crucial to unsupervised feature learning algorithms; one method embeds a kernel learning within CNNs while the other couples clustering with CNNs. We then further improved the quality of image features using domain adaptation methods (DAs) that learn representations that are invariant to domains with different data distributions. We present a deep unsupervised feature extractor to transform the feature maps from the pre-trained CNN on natural images to a set of non-redundant and relevant medical image features. Our feature extractor preserves meaningful generic features from the pre-trained domain and learns specific local features that are more representative of the medical image data. We conducted extensive experiments on 4 public datasets which have diverse visual characteristics of medical images including X-ray, dermoscopic and CT images. Our results show that our methods had better accuracy when compared to other conventional unsupervised methods and competitive accuracy to methods that used state-of-the-art supervised CNNs. Our findings suggest that our methods could scale to many different transfer learning or domain adaptation approaches where they have none or small sets of labelled data.

Binarization represents a key role in many document image analysis workflows. The current state of the art considers the use of supervised learning, and specifically deep neural networks. However, it is very difficult for the same model to work successfully in a number of document styles, since the set of potential domains is very heterogeneous. We study a multi-source domain adaptation strategy for binarization. Within this scenario, we look into a novel hypothesis where a specialized binarization model must be selected to be used over a target domain, instead of a single model that tries to generalize across multiple domains. The problem then boils down to, given several specialized models and a new target set, deciding which model to use. We propose here a simple way to address this question by using a domain classifier, that estimates which of the source models must be considered to binarize the new target domain. Our experiments on several datasets, including different text styles and music scores, show that our initial hypothesis is quite promising, yet the way to deal with the decision of which model to use still shows great room for improvement.

This work proposes a novel unsupervised cross-modality adaptive segmentation method for medical images to tackle the performance degradation caused by the severe domain shift when neural networks are being deployed to unseen modalities. The proposed method is an end-2-end framework, which conducts appearance transformation via a domain-shared shallow content encoder and two domain-specific decoders. The feature extracted from the encoder is enhanced to be more domain-invariant by a similarity learning task using the proposed Semantic Similarity Mining (SSM) module which has a strong help of domain adaptation. The domain-invariant latent feature is then fused into the target domain segmentation sub-network, trained using the original target domain images and the translated target images from the source domain in the framework of adversarial training. The adversarial training is effective to narrow the remaining gap between domains in semantic space after appearance alignment. Experimental results on two challenging datasets demonstrate that our method outperforms the state-of-the-art approaches.