Search

Domain adaptation techniques learn transferable knowledge from a source domain to a target domain and train models that generalize well in the target domain. Unfortunately, a majority of the existing techniques are only applicable to scenarios that the target-domain data in the task of interest is available for training, yet this is not often true in practice. In general, human beings are experts in generalization across domains. For example, a baby can easily identify the bear from a clipart image after learning this category of animal from the photo images. To reduce the gap between the generalization ability of human and that of machines, we propose a new solution to the challenging zero-shot domain adaptation (ZSDA) problem, where only a single source domain is available and the target domain for the task of interest is not accessible. Inspired by the observation that the knowledge about domain correlation can improve our generalization ability, we explore the correlation between source domain and target domain in an irrelevant knowledge task ([Formula: see text]-task), where dual-domain samples are available. We denote the task of interest as the question task ([Formula: see text]-task) and synthesize its non-accessible target-domain as such that these two tasks have the shared domain correlation. In order to realize our idea, we introduce a new network structure, i.e., conditional coupled generative adversarial networks (CoCoGAN), by extending the coupled generative adversarial networks (CoGAN) into a conditioning model. With a pair of coupling GANs, our CoCoGAN is able to capture the joint distribution of data samples across two domains and two tasks. For CoCoGAN training in a ZSDA task, we introduce three supervisory signals, i.e., semantic relationship consistency across domains, global representation alignment across tasks, and alignment consistency across domains. Experimental results demonstrate that our method can learn a suitable model for the non-accessible target domain and outperforms the existing state of the arts in both image classification and semantic segmentation.

Background: As a physiological signal, EEG data cannot be subjectively changed or hidden. Compared with other physiological signals, EEG signals are directly related to human cortical activities with excellent temporal resolution. After the rapid development of machine learning and artificial intelligence, the analysis and calculation of EEGs has made great progress, leading to a significant boost in performances for content understanding and pattern recognition of brain activities across the areas of both neural science and computer vision. While such an enormous advance has attracted wide range of interests among relevant research communities, EEG-based classification of brain activities evoked by images still demands efforts for further improvement with respect to its accuracy, generalization, and interpretation, yet some characters of human brains have been relatively unexplored., Methods: We propose a region-level stacked bi-directional deep learning framework for EEG-based image classification. Inspired by the hemispheric lateralization of human brains, we propose to extract additional information at regional level to strengthen and emphasize the differences between two hemispheres. The stacked bi-directional long short-term memories are used to capture the dynamic correlations hidden from both the past and the future to the current state in EEG sequences., Results: Extensive experiments are carried out and our results demonstrate the effectiveness of our proposed framework. Compared with the existing state-of-the-arts, our framework achieves outstanding performances in EEG-based classification of brain activities evoked by images. In addition, we find that the signals of Gamma band are not only useful for achieving good performances for EEG-based image classification, but also play a significant role in capturing relationships between the neural activations and the specific emotional states., Conclusions: Our proposed framework provides an improved solution for the problem that, given an image used to stimulate brain activities, we should be able to identify which class the stimuli image comes from by analyzing the EEG signals. The region-level information is extracted to preserve and emphasize the hemispheric lateralization for neural functions or cognitive processes of human brains. Further, stacked bi-directional LSTMs are used to capture the dynamic correlations hidden in EEG data. Extensive experiments on standard EEG-based image classification dataset validate that our framework outperforms the existing state-of-the-arts under various contexts and experimental setups.

The logistic regression is a widely used method for multimedia classification. However, when it is applied to high-order data such as video sequences, traditional vector-based logistic regression often incurs loss of space-time structural information. The tensor extension method based on CP (CANDECOMP/PARAFAC) decomposition is powerful for capturing the multilinear latent information. The existing CP algorithms require the tensor rank to be manually specified, however, the determination of tensor rank remains a challenging problem especially for CP rank. To effectively exploit underlying space-time structural in video sequences, we propose a tensor-based logistic regression learning algorithm, in which the weight parameter are regarded to be a tensor, calculated after the CP tensor decomposition. We introduce a regularization term, L(2,1)-norm, into the logistic tensor regression, and automatically select the CP rank, making it adaptive to the input videos for improved weight tensor and thus classification performances. Extensive experimental results in comparison with five state-of-the-art regression methods support that our proposed algorithm achieves the best classification performances, providing a good potential for a range of applications towards computerized video classifications via tensor-based video descriptions. [ABSTRACT FROM AUTHOR]

Nowadays, facial expression synthesis is widely used in expression simulation, recognition and animation. While a variety of works available in literature are 3D based, these algorithms usually require face matching and trimming, are thus time consuming. In this work, an automatic algorithm for expression synthesis in 2D space is proposed, which mainly consists of three stages.The optimum matching of three sets of feature points on the faces of source neutral (Fsn), source expression (Fse) and target neutral (Ftn) are obtained in the first stage. Different components on the target face are deformed by learning from not only the displacements but also the geometry shape of face Fse in the second stage. In the last stage, the details of the source expression are mapped onto the corresponding positions on the target face by fitting of the lighting differences. Experimental results on expression synthesis with large geometry deformation and lighting difference show that the proposed algorithm is able to accurately preserve the geometry deformation, and the synthesized expressions are visually realistic. [ABSTRACT FROM AUTHOR]

Abstract: Pattern recognition using Dynamic Bayesian Networks (DBNs) is currently a growing area of study. The classification performance greatly relies on the choice of a DBN model that will best describe the dependencies in each class of data. In this paper, we present DBN models trained for the classification of handwritten digit. Two approaches to improve the suitability of the models are presented. One uses a fixed DBN structure, and is based on an Evolutionary Algorithm optimisation of the selection and of the layout of the observations for each class of data. The second approach is about learning part of the structure of the models from the training set of each class. Parameter learning is then performed for each DBN. Classification results are presented for the described models, and compared with previously published results. Both approaches were found to improve the recognition rate compared to previous results. [Copyright &y& Elsevier]

Abstract: In this paper, we describe a novel multiclass boosting algorithm, EDBoost, to achieve robust face recognition directly in JPEG compressed domain. In comparison with existing boosting algorithms, the proposed EDBoost exploits Euclidean distance (ED) to eliminate non-effective weak classifiers in each iteration of the boosted learning, and hence improves both feature selection and classifier learning by using fewer weak classifiers and producing lower error rates. When applied to face recognition, the EDBoost algorithm is capable of selecting the most discriminative DCT features directly in JPEG compressed domain to achieve high recognition performances. In addition, a new DC replacement scheme is also proposed to reduce the effect of illumination changes. In comparison with the existing techniques, the proposed scheme achieves robust face recognition without losing the important information carried by all DC coefficients. Extensive experiments support the conclusion that the proposed algorithm outperforms all representative existing techniques in terms of boosted learning, multiclass classification, lighting effect reduction and face recognition rates. [Copyright &y& Elsevier]

Manifold learning is an effective dimension reduction method to extract nonlinear structures from high dimensional data. Recently, manifold learning has received much attention within the research communities of image analysis, computer vision and document data analysis. In this paper, we propose a boosted manifold learning algorithm towards automatic 2D face recognition by using AdaBoost to select the best possible discriminating projection for manifold learning to exploit the strength of both techniques. Experimental results support that the proposed algorithm improves over existing benchmarks in terms of stability and recognition precision rates. [ABSTRACT FROM AUTHOR]

Abstract: Video temporal segmentation is normally the first and important step for content-based video applications. Many features including the pixel difference, colour histogram, motion, and edge information etc. have been widely used and reported in the literature to detect shot cuts inside videos. Although existing research on shot cut detection is active and extensive, it still remains a challenge to achieve accurate detection of all types of shot boundaries with one single algorithm. In this paper, we propose a fuzzy logic approach to integrate hybrid features for detecting shot boundaries inside general videos. The fuzzy logic approach contains two processing modes, where one is dedicated to detection of abrupt shot cuts including those short dissolved shots, and the other for detection of gradual shot cuts. These two modes are unified by a mode-selector to decide which mode the scheme should work on in order to achieve the best possible detection performances. By using the publicly available test data set from Carleton University, extensive experiments were carried out and the test results illustrate that the proposed algorithm outperforms the representative existing algorithms in terms of the precision and recall rates. [Copyright &y& Elsevier]

Abstract: In this paper, we describe a combined approach with fuzzy logic and structure tensor towards improved enhancement of possible MCs (microcalcifications) in digital mammograms. The proposed contrast enhancement algorithm has two operational components. One is structure tensor operator and the other is fuzzy enhancement operator, both of which are arranged in parallel to process the input digital mammograms. While the structure tensor operator processes the digital mammograms and produces a corresponding eigen-image to highlight the region-of-interests, the fuzzy enhancement operator fuzzifies the mammogram via the maximum fuzzy entropy principle in fuzzy domain. As a result, the local fuzzy contrast can be extracted and modified adaptively in accordance with the information provided by the eigen-image, and those non-MCs regions are suppressed by being taken as noise. After that, the mammogram is transformed back to the pixel domain and the enhanced mammogram is constructed. Extensive experimental results show that our proposed algorithm outperforms the existing benchmark in terms of cost figures across the whole range of true positive fractions (TPF). [Copyright &y& Elsevier]

Tensors representations are widely used in multimedia applications. As a key step of tensor processing, the rank-1 tensor decomposition (i.e., the CANDECOMP/PARAFAC (CP) decomposition) always requires the estimation of the tensor rank. The ℓ 2 , 1 -norm has been shown to be effective for tensor rank selection. The existing tensor rank selection algorithm force the same columns of the tensor matrices to simultaneously become zero. However, the real sparse columns for different factor matrices may be different. Such strategy does not really uncover the sparse information of each factor matrix. In this paper, we add a separable ℓ 2 , 1 -norm on multiple factor matrices to obtain real sparse results along to different modes. And then different sparse results are assembled into a joint sparse pattern for tensor rank selection. This added separable regularization term has twofold role in enhancing the effect of regularization for each factor matrix and fully utilizing the knowledge of multiple factor matrices to facilitate decision making. In order to effectively exploit the structure information of multimedia data, we propose a model of tensor bag of words (tBOW) as the direct input of our algorithms. In the experiments, we apply the proposed algorithms to three representative tasks of multimedia analysis, i.e., image classification, video action recognition, and head pose estimation. Experimental results on three open benchmark datasets show that our algorithms are effective to multimedia analysis. [ABSTRACT FROM AUTHOR]

Abstract: Recognition of handwritten Arabic cursive texts is a complex task due to the similarities between letters under different writing styles. In this paper, a word-based off-line recognition system is proposed, using Hidden Markov Models (HMMs). The method employed involves three stages, namely preprocessing, feature extraction and classification. First, words from input scripts are segmented and normalized. Then, a set of intensity features are extracted from each of the segmented words, which is based on a sliding window moving across each mirrored word image. Meanwhile, structure-like features are also extracted including number of subwords and diacritical marks. Finally, these features are applied in a combined scheme for classification. Intensity features are used to train a HMM classifier, whose results are re-ranked using structure-like features for improved recognition rate. In order to validate the proposed techniques, extensive experiments were carried out using the IFN/ENIT database which contains 32,492 handwritten Arabic words. The proposed algorithm yields superior results of improved accuracy in comparison with several typical methods. [Copyright &y& Elsevier]

This paper proposes a video super-resolution method based on an adaptive superpixel-guided auto-regressive (AR) model. Key-frames are automatically selected and super-resolved by a sparse regression method. Non-key-frames are super-resolved by exploiting the spatio-temporal correlations: the temporal correlation is exploited by an optical flow method while the spatial correlation is modeled by a superpixel-guided AR model. Experimental results show that the proposed method outperforms state-of-the-art methods in terms of both subjective visual quality and objective peak signal-to-noise ratio (PSNR). The proposed method requires less computation and is suitable for practical applications. [ABSTRACT FROM AUTHOR]