Search

Food recognition has captured numerous research attention for its importance for health-related applications. The existing approaches mostly focus on the categorization of food according to dish names, while ignoring the underlying ingredient composition. In reality, two dishes with the same name do not necessarily share the exact list of ingredients. Therefore, the dishes under the same food category are not mandatorily equal in nutrition content. Nevertheless, due to limited datasets available with ingredient labels, the problem of ingredient recognition is often overlooked. Furthermore, as the number of ingredients is expected to be much less than the number of food categories, ingredient recognition is more tractable in the real-world scenario. This paper provides an insightful analysis of three compelling issues in ingredient recognition. These issues involve recognition in either image-level or region level, pooling in either single or multiple image scales, learning in either single or multi-task manner. The analysis is conducted on a large food dataset, Vireo Food-251, contributed by this paper. The dataset is composed of 169,673 images with 251 popular Chinese food and 406 ingredients. The dataset includes adequate challenges in scale and complexity to reveal the limit of the current approaches in ingredient recognition.

Face hallucination or super-resolution is a practical application of general image super-resolution which has been recently studied by many researchers. The challenge of good face hallucination comes from a variety of poses, illuminations, facial expressions, and other degradations. In many proposed methods, researchers resolve it by using a generative neural network to reduce the perceptual loss so we can generate a photo-realistic image. The problem is that researchers usually overlook the fidelity of the super-resolved image which could affect further facial image processing. Meanwhile, many CNN based approaches cascade multiple networks to extract facial prior information to improve super-resolution quality. Because of the end-to-end design, the details are missing for investigation. In this paper, we combine new techniques in convolutional neural network and random forests to a Hierarchical CNN based Random Forests (HCRF) approach for face super-resolution in a coarse-to-fine manner. In the proposed approach, we focus on a general approach that can handle facial images with various conditions without pre-processing. To the best of our knowledge, this is the first paper that combines the advantages of deep learning with random forests for face super-resolution. To achieve superior performance, we propose two novel CNN models for coarse facial image super-resolution and segmentation and then apply new random forests to target on local facial features refinement making use of the segmentation results. Extensive benchmark experiments on subjective and objective evaluation show that HCRF can achieve comparable speed and competitive performance compared with state-of-the-art super-resolution approaches for very low-resolution images.

Recently, food recognition has received more and more attention in image processing and computer vision for its great potential applications in human health. Most of the existing methods directly extracted deep visual features via convolutional neural networks (CNNs) for food recognition. Such methods ignore the characteristics of food images and are, thus, hard to achieve optimal recognition performance. In contrast to general object recognition, food images typically do not exhibit distinctive spatial arrangement and common semantic patterns. In this paper, we propose a multi-scale multi-view feature aggregation (MSMVFA) scheme for food recognition. MSMVFA can aggregate high-level semantic features, mid-level attribute features, and deep visual features into a unified representation. These three types of features describe the food image from different granularity. Therefore, the aggregated features can capture the semantics of food images with the greatest probability. For that solution, we utilize additional ingredient knowledge to obtain mid-level attribute representation via ingredient-supervised CNNs. High-level semantic features and deep visual features are extracted from class-supervised CNNs. Considering food images do not exhibit distinctive spatial layout in many cases, MSMVFA fuses multi-scale CNN activations for each type of features to make aggregated features more discriminative and invariable to geometrical deformation. Finally, the aggregated features are more robust, comprehensive, and discriminative via two-level fusion, namely multi-scale fusion for each type of features and multi-view aggregation for different types of features. In addition, MSMVFA is general and different deep networks can be easily applied into this scheme. Extensive experiments and evaluations demonstrate that our method achieves state-of-the-art recognition performance on three popular large-scale food benchmark datasets in Top-1 recognition accuracy. Furthermore, we expect this paper will further the agenda of food recognition in the community of image processing and computer vision.

Integer Linear Programming (ILP) formulations of multi-label Markov random fields (MRFs) models with global connectivity priors were investigated previously in computer vision. In these works, only Linear Programming (LP) relaxations [1] or simplified versions [2] of the problem were solved. This paper investigates the ILP of MRF with exact connectivity priors via a branch-and-cut method, which provably finds globally optimal solutions. It enforces connectivity priors iteratively by a cutting plane method, and provides feasible solutions with a guarantee on sub-optimality even if we terminate it earlier. The proposed ILP can be applied as a post-processing method on top of any existing multi-label segmentation approach. As it provides globally optimal solution, it can be used off-line to serve as quality check for any fast on-line algorithm. Furthermore, the scribble based model presented in this paper could be potentially used to generate ground-truth proposals for any deep learning based segmentation. We demonstrate the power and usefulness of our model by extensive experiments on the BSDS500 and PASCAL VOC dataset. The experiments show that our proposed model achieves great performance, yielding provably global optimum in most instances and that provably good optimization solutions also provide good segmentation accuracy, even with the limited computing time of few seconds.

Single image dehazing has been a challenging problem because of its ill-posed nature. For this reason, numerous efforts have been made in the field of haze removal. This paper proposes a simple, fast, and powerful algorithm for haze removal. The medium transmission is derived as a function of the saturation of the scene radiance only, and the saturation of scene radiance is estimated using a simple stretching method. A different medium transmission can be estimated for each pixel because this method does not assume that transmission is constant in a small patch. Furthermore, this paper presents a color veil removing algorithm, which is useful for an image with fine or yellow dust, using the white balance technique. The proposed algorithm requires no training, prior, and refinement process. The simulation results show that the proposed dehazing scheme outperforms state-of-the-art dehazing approaches in terms of both computational complexity and dehazing efficiency.

Recent research have shown the potential of using convolutional neural networks (CNNs) to accomplish single image dehazing. In this paper, we take one step further to explore the possibility of exploiting a network to perform haze removal for videos. Unlike single image dehazing, video-based approaches can take advantage of the abundant information that exists across neighboring frames. In this paper, assuming that a scene point yields highly correlated transmission values between adjacent video frames, we develop a deep learning solution for video dehazing, where a CNN is trained end-to-end to learn how to accumulate information across frames for transmission estimation. The estimated transmission map is subsequently used to recover a haze-free frame via atmospheric scattering model. In addition, as the semantic information of a scene provides a strong prior for image restoration, we propose to incorporate global semantic priors as input to regularize the transmission maps so that the estimated maps can be smooth in the regions of the same object and only discontinuous across the boundaries of different objects. To train this network, we generate a dataset consisted of synthetic hazy and haze-free videos for supervision based on the NYU depth dataset. We show that the features learned from this dataset are capable of removing haze that arises in outdoor scenes in a wide range of videos. Extensive experiments demonstrate that the proposed algorithm performs favorably against the state-of-the-art methods on both synthetic and real-world videos.

Different from the object motion blur, the defocus blur is caused by the limitation of the cameras' depth of field. The defocus amount can be characterized by the parameter of point spread function and thus forms a defocus map. In this paper, we propose a new network architecture called Defocus Image Deblurring Auxiliary Learning Net (DID-ANet), which is specifically designed for single image defocus deblurring by using defocus map estimation as auxiliary task to improve the deblurring result. To facilitate the training of the network, we build a novel and large-scale dataset for single image defocus deblurring, which contains the defocus images, the defocus maps and the all-sharp images. To the best of our knowledge, the new dataset is the first large-scale defocus deblurring dataset for training deep networks. Moreover, the experimental results demonstrate that the proposed DID-ANet outperforms the state-of-the-art methods for both tasks of defocus image deblurring and defocus map estimation, both quantitatively and qualitatively. The dataset, code, and model is available on GitHub: https://github.com/xytmhy/DID-ANet-Defocus-Deblurring.

Ultra-high definition (UHD) 360 videos encoded in fine quality are typically too large to stream in its entirety over bandwidth (BW)-constrained networks. One popular approach is to interactively extract and send a spatial sub-region corresponding to a viewer's current field-of-view (FoV) in a head-mounted display (HMD) for more BW-efficient streaming. Due to the non-negligible round-trip-time (RTT) delay between server and client, accurate head movement prediction foretelling a viewer's future FoVs is essential. In this paper, we cast the head movement prediction task as a sparse directed graph learning problem: three sources of relevant information-collected viewers' head movement traces, a 360 image saliency map, and a biological human head model-are distilled into a view transition Markov model. Specifically, we formulate a constrained maximum a posteriori (MAP) problem with likelihood and prior terms defined using the three information sources. We solve the MAP problem alternately using a hybrid iterative reweighted least square (IRLS) and Frank-Wolfe (FW) optimization strategy. In each FW iteration, a linear program (LP) is solved, whose runtime is reduced thanks to warm start initialization. Having estimated a Markov model from data, we employ it to optimize a tile-based 360 video streaming system. Extensive experiments show that our head movement prediction scheme noticeably outperformed existing proposals, and our optimized tile-based streaming scheme outperformed competitors in rate-distortion performance.

RGBT tracking has attracted increasing attention since RGB and thermal infrared data have strong complementary advantages, which could make trackers all-day and all-weather work. However, how to effectively represent RGBT data for visual tracking remains unstudied well. Existing works usually focus on extracting modality-shared or modality-specific information, but the potentials of these two cues are not well explored and exploited in RGBT tracking. In this paper, we propose a novel multi-adapter network to jointly perform modality-shared, modality-specific and instance-aware target representation learning for RGBT tracking. To this end, we design three kinds of adapters within an end-to-end deep learning framework. In specific, we use the modified VGG-M as the generality adapter to extract the modality-shared target representations.To extract the modality-specific features while reducing the computational complexity, we design a modality adapter, which adds a small block to the generality adapter in each layer and each modality in a parallel manner. Such a design could learn multilevel modality-specific representations with a modest number of parameters as the vast majority of parameters are shared with the generality adapter. We also design instance adapter to capture the appearance properties and temporal variations of a certain target. Moreover, to enhance the shared and specific features, we employ the loss of multiple kernel maximum mean discrepancy to measure the distribution divergence of different modal features and integrate it into each layer for more robust representation learning. Extensive experiments on two RGBT tracking benchmark datasets demonstrate the outstanding performance of the proposed tracker against the state-of-the-art methods., Accepted by IEEE TIP

The existing fusion-based RGB-D salient object detection methods usually adopt the bistream structure to strike a balance in the fusion trade-off between RGB and depth (D). While the D quality usually varies among the scenes, the state-of-the-art bistream approaches are depth-quality-unaware, resulting in substantial difficulties in achieving complementary fusion status between RGB and D and leading to poor fusion results for low-quality D. Thus, this paper attempts to integrate a novel depth-quality-aware subnet into the classic bistream structure in order to assess the depth quality prior to conducting the selective RGB-D fusion. Compared to the SOTA bistream methods, the major advantage of our method is its ability to lessen the importance of the low-quality, no-contribution, or even negative-contribution D regions during RGB-D fusion, achieving a much improved complementary status between RGB and D. Our source code and data are available online at https://github.com/qdu1995/DQSD .

Recently, the residual learning strategy has been integrated into the convolutional neural network (CNN) for single image super-resolution (SISR), where the CNN is trained to estimate the residual images. Recognizing that a residual image usually consists of high-frequency details and exhibits cartoon-like characteristics, in this paper, we propose a deep shearlet residual learning network (DSRLN) to estimate the residual images based on the shearlet transform. The proposed network is trained in the shearlet transform-domain which provides an optimal sparse approximation of the cartoon-like image. Specifically, to address the large statistical variation among the shearlet coefficients, a dual-path training strategy and a data weighting technique are proposed. Extensive evaluations on general natural image datasets as well as remote sensing image datasets show that the proposed DSRLN scheme achieves close results in PSNR to the state-of-the-art deep learning methods, using much less network parameters.

Despite the remarkable progress in recent years, person Re-Identification (ReID) approaches frequently fail in cases where the semantic body parts are misaligned between the detected human boxes. To mitigate such cases, we propose a novel High-Order ReID (HOReID) framework that enables semantic pose alignment by aggregating the fine-grained part details of multilevel feature maps. The HOReID adopts a high-order mapping of multilevel feature similarities in order to emphasize the differences of the similarities between aligned and misaligned part pairs in two person images. Since the similarities of misaligned part pairs are reduced, the HOReID enhances pose-robustness within the learned features. We show that our method derives from an intuitive and interpretable motivation and elegantly reduces the misalignment problem without using any prior knowledge from human pose annotations or pose estimation networks. This paper theoretically and experimentally demonstrates the effectiveness of the proposed HOReID, achieving superior performance over the state-of-the-art methods on the four large-scale person ReID datasets.

The trade-off between spatial and spectral resolution is one of the fundamental issues in hyperspectral images (HSI). Given the challenges of directly acquiring high-resolution hyperspectral images (HR-HSI), a compromised solution is to fuse a pair of images: one has high-resolution (HR) in the spatial domain but low-resolution (LR) in spectral-domain and the other vice versa. Model-based image fusion methods including pan-sharpening aim at reconstructing HR-HSI by solving manually designed objective functions. However, such hand-crafted prior often leads to inevitable performance degradation due to a lack of end-to-end optimization. Although several deep learning-based methods have been proposed for hyperspectral pan-sharpening, HR-HSI related domain knowledge has not been fully exploited, leaving room for further improvement. In this paper, we propose an iterative Hyperspectral Image Super-Resolution (HSISR) algorithm based on a deep HSI denoiser to leverage both domain knowledge likelihood and deep image prior. By taking the observation matrix of HSI into account during the end-to-end optimization, we show how to unfold an iterative HSISR algorithm into a novel model-guided deep convolutional network (MoG-DCN). The representation of the observation matrix by subnetworks also allows the unfolded deep HSISR network to work with different HSI situations, which enhances the flexibility of MoG-DCN. Extensive experimental results are reported to demonstrate that the proposed MoG-DCN outperforms several leading HSISR methods in terms of both implementation cost and visual quality. The code is available at https://see.xidian.edu.cn/faculty/wsdong/Projects/MoG-DCN.htm .

Classification remains challenging when confronted with the existence of multi-view data with limited labels. In this paper, we propose an embedding regularizer learning scheme for multi-view semi-supervised classification (ERL-MVSC). The proposed framework integrates diversity, sparsity and consensus to dexterously manipulate multi-view data with limited labels. To encourage diversity, ERL-MVSC recasts a linear regression model to derive view-specific embedding regularizers and automatically determines their weights. This is able to tactfully incorporate complementary information of different views. To ensure sparsity, ERL-MVSC imposes $\ell _{2,1}$ -norm on a fused embedding regularizer to exploit the sparse local structure of samples, thereby conveying valuable classification information and enhancing the robustness against noise/outliers. To enhance consensus, ERL-MVSC learns a shared predicted label matrix, which serves as the comment target of multi-view classification. With these techniques, we formulate ERL-MVSC as a joint optimization problem of an embedding regularizer and a predicted label matrix, which can be solved by a coordinate descent method. Extensive experimental results on real-world datasets demonstrate the effectiveness and superiority of the proposed algorithm.

Shadow removal can significantly improve the image visual quality and has many applications in computer vision. Deep learning methods based on CNNs have become the most effective approach for shadow removal by training on either paired data, where both the shadow and underlying shadow-free versions of an image are known, or unpaired data, where shadow and shadow-free training images are totally different with no correspondence. In practice, CNN training on unpaired data is more preferred given the easiness of training data collection. In this paper, we present a new Lightness-Guided Shadow Removal Network (LG-ShadowNet) for shadow removal by training on unpaired data. In this method, we first train a CNN module to compensate for the lightness and then train a second CNN module with the guidance of lightness information from the first CNN module for final shadow removal. We also introduce a loss function to further utilise the colour prior of existing data. Extensive experiments on widely used ISTD, adjusted ISTD and USR datasets demonstrate that the proposed method outperforms the state-of-the-art methods with training on unpaired data., Submitted to IEEE TIP

Modern computer vision requires processing large amounts of data, both while training the model and/or during inference, once the model is deployed. Scenarios where images are captured and processed in physically separated locations are increasingly common (e.g. autonomous vehicles, cloud computing). In addition, many devices suffer from limited resources to store or transmit data (e.g. storage space, channel capacity). In these scenarios, lossy image compression plays a crucial role to effectively increase the number of images collected under such constraints. However, lossy compression entails some undesired degradation of the data that may harm the performance of the downstream analysis task at hand, since important semantic information may be lost in the process. Moreover, we may only have compressed images at training time but are able to use original images at inference time, or vice versa, and in such a case, the downstream model suffers from covariate shift. In this paper, we analyze this phenomenon, with a special focus on vision-based perception for autonomous driving as a paradigmatic scenario. We see that loss of semantic information and covariate shift do indeed exist, resulting in a drop in performance that depends on the compression rate. In order to address the problem, we propose dataset restoration, based on image restoration with generative adversarial networks (GANs). Our method is agnostic to both the particular image compression method and the downstream task; and has the advantage of not adding additional cost to the deployed models, which is particularly important in resource-limited devices. The presented experiments focus on semantic segmentation as a challenging use case, cover a broad range of compression rates and diverse datasets, and show how our method is able to significantly alleviate the negative effects of compression on the downstream visual task., Comment: Accepted for publication in IEEE Transactions on Image Processing; 15 pages, 15 figures

RGB-D saliency detection is receiving more and more attention in recent years. There are many efforts have been devoted to this area, where most of them try to integrate the multi-modal information, i.e. RGB images and depth maps, via various fusion strategies. However, some of them ignore the inherent difference between the two modalities, which leads to the performance degradation when handling some challenging scenes. Therefore, in this paper, we propose a novel RGB-D saliency model, namely Dynamic Selective Network (DSNet), to perform salient object detection (SOD) in RGB-D images by taking full advantage of the complementarity between the two modalities. Specifically, we first deploy a cross-modal global context module (CGCM) to acquire the high-level semantic information, which can be used to roughly locate salient objects. Then, we design a dynamic selective module (DSM) to dynamically mine the cross-modal complementary information between RGB images and depth maps, and to further optimize the multi-level and multi-scale information by executing the gated and pooling based selection, respectively. Moreover, we conduct the boundary refinement to obtain high-quality saliency maps with clear boundary details. Extensive experiments on eight public RGB-D datasets show that the proposed DSNet achieves a competitive and excellent performance against the current 17 state-of-the-art RGB-D SOD models.

At present, most semantic segmentation models rely on the excellent feature extraction capabilities of a deep learning network structure. Although these models can achieve excellent performance on multiple datasets, ways of refining the target main body segmentation and overcoming the performance limitation of deep learning networks are still a research focus. We discovered a pan-class intrinsic relevance phenomenon among targets that can link the targets cross-class. This cross-class strategy is different from the latest semantic segmentation model via context where targets are divided into an intra-class and inter-class. This paper proposes a model for refining the target main body segmentation using multi-target pan-class intrinsic relevance. The main contributions of the proposed model can be summarized as follows: a) The multi-target pan-class intrinsic relevance prior knowledge establishment (RPK-Est) module builds the prior knowledge of the intrinsic relevance to lay the foundation for the following extraction of the pan-class intrinsic relevance feature. b) The multi-target pan-class intrinsic relevance feature extraction (RF-Ext) module is designed to extract the pan-class intrinsic relevance feature based on the proposed multi-target node graph and graph convolution network. c) The multi-target pan-class intrinsic relevance feature integration (RF-Int) module is proposed to integrate the intrinsic relevance features and semantic features by a generative adversarial learning strategy at the gradient level, which can make intrinsic relevance features play a role in semantic segmentation. The proposed model achieved outstanding performance in semantic segmentation testing on four authoritative datasets compared to other state-of-the-art models.

Hyperspectral imagery (HSI) contains rich spectral information, which is beneficial to many tasks. However, acquiring HSI is difficult because of the limitations of current imaging technology. As an alternative method, spectral super-resolution aims at reconstructing HSI from its corresponding RGB image. Recently, deep learning has shown its power to this task, but most of the used networks are transferred from other domains, such as spatial super-resolution. In this paper, we attempt to design a spectral super-resolution network by taking advantage of two intrinsic properties of HSI. The first one is the spectral correlation. Based on this property, a decomposition subnetwork is designed to reconstruct HSI. The other one is the projection property, i.e., RGB image can be regarded as a three-dimensional projection of HSI. Inspired from it, a self-supervised subnetwork is constructed as a constraint to the decomposition subnetwork. These two subnetworks constitute our end-to-end super-resolution network. In order to test the effectiveness of it, we conduct experiments on three widely used HSI datasets (i.e., CAVE, NUS, and NTIRE2018). Experimental results show that our proposed network can achieve competitive reconstruction performance in comparison with several state-of-the-art networks.

Defocus blur detection (DBD), which has been widely applied to various fields, aims to detect the out-of-focus or in-focus pixels from a single image. Despite the fact that the deep learning based methods applied to DBD have outperformed the hand-crafted feature based methods, the performance cannot still meet our requirement. In this paper, a novel network is established for DBD. Unlike existing methods which only learn the projection from the in-focus part to the ground-truth, both in-focus and out-of-focus pixels, which are completely and symmetrically complementary, are taken into account. Specifically, two symmetric branches are designed to jointly estimate the probability of focus and defocus pixels, respectively. Due to their complementary constraint, each layer in a branch is affected by an attention obtained from another branch, effectively learning the detailed information which may be ignored in one branch. The feature maps from these two branches are then passed through a unique fusion block to simultaneously get the two-channel output measured by a complementary loss. Additionally, instead of estimating only one binary map from a specific layer, each layer is encouraged to estimate the ground truth to guide the binary map estimation in its linked shallower layer followed by a top-to-bottom combination strategy, gradually exploiting the global and local information. Experimental results on released datasets demonstrate that our proposed method remarkably outperforms state-of-the-art algorithms.

3D object recognition is one of the most important tasks in 3D data processing, and has been extensively studied recently. Researchers have proposed various 3D recognition methods based on deep learning, among which a class of view-based approaches is a typical one. However, in the view-based methods, the commonly used view pooling layer to fuse multi-view features causes a loss of visual information. To alleviate this problem, in this paper, we construct a novel layer called Dynamic Routing Layer (DRL) by modifying the dynamic routing algorithm of capsule network, to more effectively fuse the features of each view. Concretely, in DRL, we use rearrangement and affine transformation to convert features, then leverage the modified dynamic routing algorithm to adaptively choose the converted features, instead of ignoring all but the most active feature in view pooling layer. We also illustrate that the view pooling layer is a special case of our DRL. In addition, based on DRL, we further present a Dynamic Routing Convolutional Neural Network (DRCNN) for multi-view 3D object recognition. Our experiments on three 3D benchmark datasets show that our proposed DRCNN outperforms many state-of-the-arts, which demonstrates the efficacy of our method.

Deep unfolding methods design deep neural networks as learned variations of optimization algorithms through the unrolling of their iterations. These networks have been shown to achieve faster convergence and higher accuracy than the original optimization methods. In this line of research, this paper presents novel interpretable deep recurrent neural networks (RNNs), designed by the unfolding of iterative algorithms that solve the task of sequential signal reconstruction (in particular, video reconstruction). The proposed networks are designed by accounting that video frames’ patches have a sparse representation and the temporal difference between consecutive representations is also sparse. Specifically, we design an interpretable deep RNN (coined reweighted-RNN) by unrolling the iterations of a proximal method that solves a reweighted version of the $\ell _{1}$ - $\ell _{1}$ minimization problem. Due to the underlying minimization model, our reweighted-RNN has a different thresholding function (alias, different activation function) for each hidden unit in each layer. In this way, it has higher network expressivity than existing deep unfolding RNN models. We also present the derivative $\ell _{1}$ - $\ell _{1}$ -RNN model, which is obtained by unfolding a proximal method for the $\ell _{1}$ - $\ell _{1}$ minimization problem. We apply the proposed interpretable RNNs to the task of video frame reconstruction from low-dimensional measurements, that is, sequential video frame reconstruction. The experimental results on various datasets demonstrate that the proposed deep RNNs outperform various RNN models.

Versatile Video Coding (VVC), as the latest standard, significantly improves the coding efficiency over its ancestor standard High Efficiency Video Coding (HEVC), but at the expense of sharply increased complexity. In VVC, the quad-tree plus multi-type tree (QTMT) structure of coding unit (CU) partition accounts for over 97% of the encoding time, due to the brute-force search for recursive rate-distortion (RD) optimization. Instead of the brute-force QTMT search, this paper proposes a deep learning approach to predict the QTMT-based CU partition, for drastically accelerating the encoding process of intra-mode VVC. First, we establish a large-scale database containing sufficient CU partition patterns with diverse video content, which can facilitate the data-driven VVC complexity reduction. Next, we propose a multi-stage exit CNN (MSE-CNN) model with an early-exit mechanism to determine the CU partition, in accord with the flexible QTMT structure at multiple stages. Then, we design an adaptive loss function for training the MSE-CNN model, synthesizing both the uncertain number of split modes and the target on minimized RD cost. Finally, a multi-threshold decision scheme is developed, achieving desirable trade-off between complexity and RD performance. Experimental results demonstrate that our approach can reduce the encoding time of VVC by 44.65%-66.88% with the negligible Bj{\o}ntegaard delta bit-rate (BD-BR) of 1.322%-3.188%, which significantly outperforms other state-of-the-art approaches., Comment: 14 pages, 10 figures, 7 tables. Published in IEEE Transactions on Image Processing (TIP), 2021

Unsupervised Domain Adaptation (UDA) aims to learn a classifier for the unlabeled target domain by leveraging knowledge from a labeled source domain with a different but related distribution. Many existing approaches typically learn a domain-invariant representation space by directly matching the marginal distributions of the two domains. However, they ignore exploring the underlying discriminative features of the target data and align the cross-domain discriminative features, which may lead to suboptimal performance. To tackle these two issues simultaneously, this paper presents a Joint Clustering and Discriminative Feature Alignment (JCDFA) approach for UDA, which is capable of naturally unifying the mining of discriminative features and the alignment of class-discriminative features into one single framework. Specifically, in order to mine the intrinsic discriminative information of the unlabeled target data, JCDFA jointly learns a shared encoding representation for two tasks: supervised classification of labeled source data, and discriminative clustering of unlabeled target data, where the classification of the source domain can guide the clustering learning of the target domain to locate the object category. We then conduct the cross-domain discriminative feature alignment by separately optimizing two new metrics: 1) an extended supervised contrastive learning, i.e., semi-supervised contrastive learning 2) an extended Maximum Mean Discrepancy (MMD), i.e., conditional MMD, explicitly minimizing the intra-class dispersion and maximizing the inter-class compactness. When these two procedures, i.e., discriminative features mining and alignment are integrated into one framework, they tend to benefit from each other to enhance the final performance from a cooperative learning perspective. Experiments are conducted on four real-world benchmarks ( e.g., Office-31, ImageCLEF-DA, Office-Home and VisDA-C). All the results demonstrate that our JCDFA can obtain remarkable margins over state-of-the-art domain adaptation methods. Comprehensive ablation studies also verify the importance of each key component of our proposed algorithm and the effectiveness of combining two learning strategies into a framework.

Multi-view representation learning (MvRL) aims to learn a consensus representation from diverse sources or domains to facilitate downstream tasks such as clustering, retrieval, and classification. Due to the limited representative capacity of the adopted shallow models, most existing MvRL methods may yield unsatisfactory results, especially when the labels of data are unavailable. To enjoy the representative capacity of deep learning, this paper proposes a novel multi-view unsupervised representation learning method, termed as Multi-view Laplacian Network (MvLNet), which could be the first deep version of the multi-view spectral representation learning method. Note that, such an attempt is nontrivial because simply combining Laplacian embedding (i.e., spectral representation) with neural networks will lead to trivial solutions. To solve this problem, MvLNet enforces an orthogonal constraint and reformulates it as a layer with the help of Cholesky decomposition. The orthogonal layer is stacked on the embedding network so that a common space could be learned for consensus representation. Compared with numerous recent-proposed approaches, extensive experiments on seven challenging datasets demonstrate the effectiveness of our method in three multi-view tasks including clustering, recognition, and retrieval. The source code could be found at www.pengxi.me .

General image super-resolution techniques have difficulties in recovering detailed face structures when applying to low resolution face images. Recent deep learning based methods tailored for face images have achieved improved performance by jointly trained with additional task such as face parsing and landmark prediction. However, multi-task learning requires extra manually labeled data. Besides, most of the existing works can only generate relatively low resolution face images (e.g., $128\times128$), and their applications are therefore limited. In this paper, we introduce a novel SPatial Attention Residual Network (SPARNet) built on our newly proposed Face Attention Units (FAUs) for face super-resolution. Specifically, we introduce a spatial attention mechanism to the vanilla residual blocks. This enables the convolutional layers to adaptively bootstrap features related to the key face structures and pay less attention to those less feature-rich regions. This makes the training more effective and efficient as the key face structures only account for a very small portion of the face image. Visualization of the attention maps shows that our spatial attention network can capture the key face structures well even for very low resolution faces (e.g., $16\times16$). Quantitative comparisons on various kinds of metrics (including PSNR, SSIM, identity similarity, and landmark detection) demonstrate the superiority of our method over current state-of-the-arts. We further extend SPARNet with multi-scale discriminators, named as SPARNetHD, to produce high resolution results (i.e., $512\times512$). We show that SPARNetHD trained with synthetic data cannot only produce high quality and high resolution outputs for synthetically degraded face images, but also show good generalization ability to real world low quality face images., Comment: TIP 2020. Codes are available at https://github.com/chaofengc/Face-SPARNet

Recently, deep learning based video super-resolution (SR) methods combine the convolutional neural networks (CNN) with motion compensation to estimate a high-resolution (HR) video from its low-resolution (LR) counterpart. However, most previous methods conduct downscaling motion estimation to handle large motions, which can lead to detrimental effects on the accuracy of motion estimation due to the reduction of spatial resolution. Besides, these methods usually treat different types of intermediate features equally, which lack flexibility to emphasize meaningful information for revealing the high-frequency details. In this paper, to solve above issues, we propose a deep dual attention network (DDAN), including a motion compensation network (MCNet) and a SR reconstruction network (ReconNet), to fully exploit the spatio-temporal informative features for accurate video SR. The MCNet progressively learns the optical flow representations to synthesize the motion information across adjacent frames in a pyramid fashion. To decrease the mis-registration errors caused by the optical flow based motion compensation, we extract the detail components of original LR neighboring frames as complementary information for accurate feature extraction. In the ReconNet, we implement dual attention mechanisms on a residual unit and form a residual attention unit to focus on the intermediate informative features for high-frequency details recovery. Extensive experimental results on numerous datasets demonstrate the proposed method can effectively achieve superior performance in terms of quantitative and qualitative assessments compared with state-of-the-art methods.

Albeit remarkable progress has been made to improve the accuracy and completeness of multi-view stereo (MVS), existing methods still suffer from either sparse reconstructions of low-textured surfaces or heavy computational burden. In this paper, we propose a Confidence-based Large-scale Dense Multi-view Stereo (CLD-MVS) method for high resolution imagery. Firstly, we formulate MVS as a multi-view depth estimation problem, and employ a normal-aware efficient PatchMatch stereo to estimate the initial depth and normal map for each reference view. A self-supervised deep learning method is then developed to predict the spatial confidence for multi-view depth maps, which is combined with cross-view consistency to generate the ground control points. Subsequently, a confidence-driven and boundary-aware interpolation scheme using static and dynamic guidance is adopted to synthesize dense depth and normal maps. Finally, a refinement procedure which leverages synthesized depth and normal as prior is conducted to estimate cross-view consistent surface. Experiments show that the proposed CLD-MVS method achieves high geometric completeness while preserving fine-scale details. In particular, it has ranked No. 1 on the ETH3D high-resolution MVS benchmark in terms of $F_{1}$ -score.

Semantic image segmentation aims to classify every pixel of a scene image to one of many classes. It implicitly involves object recognition, localization, and boundary delineation. In this paper, we propose a segmentation network called CGBNet to enhance the paring results by context encoding and multi-path decoding. We first propose a context encoding module that generates context contrasted local feature to make use of the informative context and the discriminative local information. This context encoding module greatly improves the segmentation performance, especially for inconspicuous objects. Furthermore, we propose a scale-selection scheme to selectively fuse the parsing results from different-scales of features at every spatial position. It adaptively selects appropriate score maps from rich scales of features. To improve the parsing results of boundary, we further propose a boundary delineation module that encourages the location-specific very-low-level feature near the boundaries to take part in the final prediction and suppresses them far from the boundaries. Without bells and whistles, the proposed segmentation network achieves very competitive performance in terms of all three different evaluation metrics consistently on the four popular scene segmentation datasets, Pascal Context, SUN-RGBD, Sift Flow, and COCO Stuff, ADE20K, and Cityscapes. Ministry of Education (MOE) This work was jointly supported by Singapore Ministry of Education Academic Research Fund (AcRF) Tier 3 Grant no: MOE2017-T3-1-001, and Zhejiang Leading Innovation Research Program 2018R01017.

Facial expression recognition, face synthesis, and face alignment are three coherently related tasks and can be solved in a joint framework. To achieve this goal, in this paper, we propose a novel end-to-end deep learning model by exploiting the expression code, geometry code and generated data jointly for simultaneous pose-invariant facial expression recognition, face image synthesis, and face alignment. The proposed deep model enjoys several merits. First, to the best of our knowledge, this is the first work to address these three tasks jointly in a unified deep model to complement and enhance each other. Second, the proposed model can effectively disentangle the global and local identity representation from different expression and geometry codes. As a result, it can automatically generate facial images with different expressions under arbitrary geometry codes. Third, these three tasks can further boost their performance for each other via our model. Extensive experimental results on three standard benchmarks demonstrate that the proposed deep model performs favorably against state-of-the-art methods on the three tasks.

Various power-constrained contrast enhancement (PCCE) techniques have been applied to an organic light emitting diode (OLED) display for reducing the power demands of the display while preserving the image quality. In this paper, we propose a new deep learning-based PCCE scheme that constrains the power consumption of the OLED displays while enhancing the contrast of the displayed image. In the proposed method, the power consumption is constrained by simply reducing the brightness a certain ratio, whereas the perceived visual quality is preserved as much as possible by enhancing the contrast of the image using a convolutional neural network (CNN). Furthermore, our CNN can learn the PCCE technique without a reference image by unsupervised learning. Experimental results show that the proposed method is superior to conventional ones in terms of image quality assessment metrics such as a visual saliency-induced index (VSI) and a measure of enhancement (EME)., Comment: Accepted to IEEE transactions on Image Processing. To be published

In this paper we address the problem of hallucinating high-resolution facial images from low-resolution inputs at high magnification factors. We approach this task with convolutional neural networks (CNNs) and propose a novel (deep) face hallucination model that incorporates identity priors into the learning procedure. The model consists of two main parts: i) a cascaded super-resolution network that upscales the low-resolution facial images, and ii) an ensemble of face recognition models that act as identity priors for the super-resolution network during training. Different from most competing super-resolution techniques that rely on a single model for upscaling (even with large magnification factors), our network uses a cascade of multiple SR models that progressively upscale the low-resolution images using steps of $2\times $ . This characteristic allows us to apply supervision signals (target appearances) at different resolutions and incorporate identity constraints at multiple-scales. The proposed C-SRIP model (Cascaded Super Resolution with Identity Priors) is able to upscale (tiny) low-resolution images captured in unconstrained conditions and produce visually convincing results for diverse low-resolution inputs. We rigorously evaluate the proposed model on the Labeled Faces in the Wild (LFW), Helen and CelebA datasets and report superior performance compared to the existing state-of-the-art.

The rapidly developing information technology leads to a fast growth of visual and textual contents, and it comes with huge challenges to make correlation and perform cross-media retrieval between images and sentences. Existing methods mainly explore cross-media correlation from either global-level instances as the whole images and sentences, or local-level fine-grained patches as the discriminative image regions and key words, which ignore the complementary information from the relation between local-level fine-grained patches. Naturally, relation understanding is highly important for learning cross-media correlation. People focus on not only the alignment between discriminative image regions and key words, but also their relations lying in the visual and textual context. Therefore, in this paper, we propose Multi-level Adaptive Visual-textual Alignment (MAVA) approach with the following contributions. First, we propose cross-media multi-pathway fine-grained network to extract not only the local fine-grained patches as discriminative image regions and key words, but also visual relations between image regions as well as textual relations from the context of sentences, which contain complementary information to exploit fine-grained characteristics within different media types. Second, we propose visual-textual bi-attention mechanism to distinguish the fine-grained information with different saliency from both local and relation levels, which can provide more discriminative hints for correlation learning. Third, we propose cross-media multi-level adaptive alignment to explore global, local and relation alignments. An adaptive alignment strategy is further proposed to enhance the matched pairs of different media types, and discard those misalignments adaptively to learn more precise cross-media correlation. Extensive experiments are conducted to perform image-sentence matching on 2 widely-used cross-media datasets, namely Flickr-30K and MS-COCO, comparing with 10 state-of-the-art methods, which can fully verify the effectiveness of our proposed MAVA approach.

Audio information has been bypassed by most of current visual attention prediction studies. However, sound could have influence on visual attention and such influence has been widely investigated and proofed by many psychological studies. In this paper, we propose a novel multi-modal saliency (MMS) model for videos containing scenes with high audio-visual correspondence. In such scenes, humans tend to be attracted by the sound sources and it is also possible to localize the sound sources via cross-modal analysis. Specifically, we first detect the spatial and temporal saliency maps from the visual modality by using a novel free energy principle. Then we propose to detect the audio saliency map from both audio and visual modalities by localizing the moving-sounding objects using cross-modal kernel canonical correlation analysis, which is first of its kind in the literature. Finally we propose a new two-stage adaptive audiovisual saliency fusion method to integrate the spatial, temporal and audio saliency maps to our audio-visual saliency map. The proposed MMS model has captured the influence of audio, which is not considered in the latest deep learning based saliency models. To take advantages of both deep saliency modeling and audio-visual saliency modeling, we propose to combine deep saliency models and the MMS model via a later fusion, and we find that an average of 5% performance gain is obtained. Experimental results on audio-visual attention databases show that the introduced models incorporating audio cues have significant superiority over state-of-the-art image and video saliency models which utilize a single visual modality.

Phase unwrapping is an ill-posed classical problem in many practical applications of significance such as 3D profiling through fringe projection, synthetic aperture radar and magnetic resonance imaging. Conventional phase unwrapping techniques estimate the phase either by integrating through the confined path (referred to as path-following methods) or by minimizing the energy function between the wrapped phase and the approximated true phase (referred to as minimum-norm approaches). However, these conventional methods have some critical challenges like error accumulation and high computational time and often fail under low SNR conditions. To address these problems, this paper proposes a novel deep learning framework for unwrapping the phase and is referred to as “PhaseNet 2.0”. The phase unwrapping problem is formulated as a dense classification problem and a fully convolutional DenseNet based neural network is trained to predict the wrap-count at each pixel from the wrapped phase maps. To train this network, we simulate arbitrary shapes and propose new loss function that integrates the residues by minimizing the difference of gradients and also uses $L_{1}$ loss to overcome class imbalance problem. The proposed method, unlike our previous approach PhaseNet, does not require post-processing, highly robust to noise, accurately unwraps the phase even at the severe noise level of −5 dB, and can unwrap the phase maps even at relatively high dynamic ranges. Simulation results from the proposed framework are compared with different classes of existing phase unwrapping methods for varying SNR values and discontinuity, and these evaluations demonstrate the advantages of the proposed framework. We also demonstrate the generality of the proposed method on 3D reconstruction of synthetic CAD models that have diverse structures and finer geometric variations. Finally, the proposed method is applied to real-data for 3D profiling of objects using fringe projection technique and digital holographic interferometry. The proposed framework achieves significant improvements over existing methods while being highly efficient with interactive frame-rates on modern GPUs.

Various weather conditions, such as rain, haze, or snow, can degrade visual quality in images/videos, which may significantly degrade the performance of related applications. In this paper, a novel framework based on sequential dual attention deep network is proposed for removing rain streaks (deraining) in a single image, called by SSDRNet (Sequential dual attention-based Single image DeRaining deep Network). Since the inherent correlation among rain steaks within an image should be stronger than that between the rain streaks and the background (non-rain) pixels, a two-stage learning strategy is implemented to better capture the distribution of rain streaks within a rainy image. The two-stage deep neural network primarily involves three blocks: residual dense blocks (RDBs), sequential dual attention blocks (SDABs), and multi-scale feature aggregation modules (MAMs), which are all delicately and specifically designed for rain removal. The two-stage strategy successfully learns very fine details of the rain steaks of the image and then clearly removes them. Extensive experimental results have shown that the proposed deep framework achieves the best performance on qualitative and quantitative metrics compared with state-of-the-art methods. The corresponding code and the trained model of the proposed SSDRNet have been available online at https://github.com/fityanul/SDAN-for-Rain-Removal .

With the rapid development of deep learning technology and other powerful tools, 3D object detection has made great progress and become one of the fastest growing field in computer vision. Many automated applications such as robotic navigation, autonomous driving, and virtual or augmented reality system require estimation of accurate 3D object location and detection. Under this requirement, many methods have been proposed to improve the performance of 3D object localization and detection. Despite recent efforts, 3D object detection is still a very challenging task due to occlusion, viewpoint variations, scale changes, and limited information in 3D scenes. In this paper, we present a comprehensive review of recent state-of-the-art approaches in 3D object detection technology. We start with some basic concepts, then describe some of the available datasets that are designed to facilitate the performance evaluation of 3D object detection algorithms. Next, we will review the state-of-the-art technologies in this area, highlighting their contributions, importance, and limitations as a guide for future research. Finally, we provide a quantitative comparison of the results of the state-of-the-art methods on the popular public datasets.

Local feature descriptor learning aims to represent distinctive images or patches with the same local features, where their representation is invariant under different types of deformation. Recent studies have demonstrated that descriptor learning based on Convolutional Neural Network (CNN) is able to improve the matching performance significantly. However, they tend to ignore the importance of sample selection during the training process, leading to unstable quality of descriptors and learning efficiency. In this paper, a dual hard batch construction method is proposed to sample the hard matching and non-matching examples for training, improving the performance of the descriptor learning on different tasks. To construct the dual hard training batches, the matching examples with the minimum similarity are selected as the hard positive pairs. For each positive pair, the most similar non-matching example is then sampled from the generated hard positive pairs in the same batch as the corresponding negative. By sampling the hard positive pairs and the corresponding hard negatives, the hard batches are produced to force the CNN model to learn the descriptors with more efforts. In addition, based on the above dual hard batch construction, an $\ell _{2}^{2}$ triplet loss function is built for optimizing the training model. Specifically, we analyze the superiority of the $\ell _{2}^{2}$ loss function when dealing with hard examples, and also demonstrate it in the experiments. With the benefits of the proposed sampling strategy and the $\ell _{2}^{2}$ triplet loss function, our method achieves better performance compared to state-of-the-art on the reference benchmarks for different matching tasks.

Semantic segmentation is a key step in scene understanding for autonomous driving. Although deep learning has significantly improved the segmentation accuracy, current high-quality models such as PSPNet and DeepLabV3 are inefficient given their complex architectures and reliance on multi-scale inputs. Thus, it is difficult to apply them to real-time or practical applications. On the other hand, existing real-time methods cannot yet produce satisfactory results on small objects such as traffic lights, which are imperative to safe autonomous driving. In this paper, we improve the performance of real-time semantic segmentation from two perspectives, methodology and data. Specifically, we propose a real-time segmentation model coined Narrow Deep Network (NDNet) and build a synthetic dataset by inserting additional small objects into the training images. The proposed method achieves 65.7% mean intersection over union (mIoU) on the Cityscapes test set with only 8.4G floating-point operations (FLOPs) on $1024\times 2048$ inputs. Furthermore, by re-training the existing PSPNet and DeepLabV3 models on our synthetic dataset, we obtained an average 2% mIoU improvement on small objects.

Non-local self similarity (NSS) is a powerful prior of natural images for image denoising. Most of existing denoising methods employ similar patches, which is a patch-level NSS prior. In this paper, we take one step forward by introducing a pixel-level NSS prior, i.e., searching similar pixels across a non-local region. This is motivated by the fact that finding closely similar pixels is more feasible than similar patches in natural images, which can be used to enhance image denoising performance. With the introduced pixel-level NSS prior, we propose an accurate noise level estimation method, and then develop a blind image denoising method based on the lifting Haar transform and Wiener filtering techniques. Experiments on benchmark datasets demonstrate that, the proposed method achieves much better performance than previous non-deep methods, and is still competitive with existing state-of-the-art deep learning based methods on real-world image denoising. The code is publicly available at https://github.com/njusthyk1972/NLH., Comment: 14 pages, 9 figures, 10 tables, accept by IEEE TIP

Due to hardware limitations of the imaging sensors, it is challenging to acquire images of high resolution in both spatial and spectral domains. Fusing a low-resolution hyperspectral image (LR-HSI) and a high-resolution multispectral image (HR-MSI) to obtain an HR-HSI in an unsupervised manner has drawn considerable attention. Though effective, most existing fusion methods are limited due to the use of linear parametric modeling for the spectral mixture process, and even the deep learning-based methods only focus on deterministic fully-connected networks without exploiting the spatial correlation and local spectral structures of the images. In this paper, we propose a novel variational probabilistic autoencoder framework implemented by convolutional neural networks, in order to fuse the spatial and spectral information contained in the LR-HSI and HR-MSI, called FusionNet. The FusionNet consists of a spectral generative network, a spatial-dependent prior network, and a spatial-spectral variational inference network, which are jointly optimized in an unsupervised manner, leading to an end-to-end fusion system. Further, for fast adaptation to different observation scenes, we give a meta-learning explanation to the fusion problem, and combine the FusionNet with meta-learning in a synergistic manner. Effectiveness and efficiency of the proposed method are evaluated based on several publicly available datasets, demonstrating that the proposed FusionNet outperforms the state-of-the-art fusion methods.

A new multi-scale deep learning (MDL) framework is proposed and exploited for conducting image interpolation in this paper. The core of the framework is a seeding network that needs to be designed for the targeted task. For image interpolation, a novel attention-aware inception network (AIN) is developed as the seeding network; it has two key stages: 1) feature extraction based on the low-resolution input image; and 2) feature-to-image mapping to enlarge image’s size or resolution. Note that the designed seeding network, AIN, needs to be trained with a matched training dataset at each scale. For that, multi-scale image patches are generated using our proposed pyramid cut , which outperforms the conventional image pyramid method by completely avoiding aliasing issue. After training, the trained AINs are then combined for processing the input image in the testing stage. Extensive experimental simulation results obtained from seven image datasets (comprising 359 images in total) have clearly shown that the proposed MAIN consistently delivers highly accurate interpolated images.

Video-based person re-identification (re-ID) refers to matching people across camera views from arbitrary unaligned video footages. Existing methods rely on supervision signals to optimise a projected space under which the distances between inter/intra-videos are maximised/minimised. However, this demands exhaustively labelling people across camera views, rendering them unable to be scaled in large networked cameras. Also, it is noticed that learning effective video representations with view invariance is not explicitly addressed for which features exhibit different distributions otherwise. Thus, matching videos for person re-ID demands flexible models to capture the dynamics in time-series observations and learn view-invariant representations with access to limited labeled training samples. In this paper, we propose a novel few-shot deep learning approach to video-based person re-ID, to learn comparable representations that are discriminative and view-invariant. The proposed method is developed on the variational recurrent neural networks (VRNNs) and trained adversarially to produce latent variables with temporal dependencies that are highly discriminative yet view-invariant in matching persons. Through extensive experiments conducted on three benchmark datasets, we empirically show the capability of our method in creating view-invariant temporal features and state-of-the-art performance achieved by our method., Comment: Appearing at IEEE Transactions on Image Processing

Multi-atlas parcellation (MAP) is carried out on a brain image by propagating and fusing labelled regions from brain atlases. Typical nonlinear registration-based label propagation is time-consuming and sensitive to inter-subject differences. Recently, deep learning parcellation (DLP) has been proposed to avoid nonlinear registration for better efficiency and robustness than MAP. However, most existing DLP methods neglect using brain atlases, which contain high-level information (e.g., manually labelled brain regions), to provide auxiliary features for improving the parcellation accuracy. In this paper, we propose a novel multi-atlas DLP method for brain parcellation. Our method is based on fully convolutional networks (FCN) and squeeze-and-excitation (SE) modules. It can automatically and adaptively select features from the most relevant brain atlases to guide parcellation. Moreover, our method is trained via a generative adversarial network (GAN), where a convolutional neural network (CNN) with multi-scale $l_{1}$ loss is used as the discriminator. Benefiting from brain atlases, our method outperforms MAP and state-of-the-art DLP methods on two public image datasets (LPBA40 and NIREP-NA0).

Sparse representation based classification (SRC), nuclear-norm matrix regression (NMR), and deep learning (DL) have achieved a great success in face recognition (FR). However, there still exist some intrinsic limitations among them. SRC and NMR based coding methods belong to one-step model, such that the latent discriminative information of the coding error vector cannot be fully exploited. DL, as a multi-step model, can learn powerful representation, but relies on large-scale data and computation resources for numerous parameters training with complicated back-propagation. Straightforward training of deep neural networks from scratch on small-scale data is almost infeasible. Therefore, in order to develop efficient algorithms that are specifically adapted for small-scale data, we propose to derive the deep models of SRC and NMR. Specifically, in this paper, we propose an end-to-end deep cascade model (DCM) based on SRC and NMR with hierarchical learning, nonlinear transformation and multi-layer structure for corrupted face recognition. The contributions include four aspects. First, an end-to-end deep cascade model for small-scale data without back-propagation is proposed. Second, a multi-level pyramid structure is integrated for local feature representation. Third, for introducing nonlinear transformation in layer-wise learning, softmax vector coding of the errors with class discrimination is proposed. Fourth, the existing representation methods can be easily integrated into our DCM framework. Experiments on a number of small-scale benchmark FR datasets demonstrate the superiority of the proposed model over state-of-the-art counterparts. Additionally, a perspective that deep-layered learning does not have to be convolutional neural network with back-propagation optimization is consolidated. The demo code is available in https://github.com/liuji93/DCM