Continuous sign language recognition (CSLR) technology enables social communication for the hearing-impaired by converting sign language videos into text. However, due to the limited receptive fields of convolutional networks and inefficient long-range dependency modeling in temporal modules, current methods find it difficult to capture cross-regional and high-order dynamic semantics in complex gestures. To address these limitations, we propose a dynamic spatiotemporal hypergraph network named HyperSign, which optimizes feature learning through innovative graph architectures. For single-frame spatial modeling, we propose a saliency-aware spatial graph construction strategy that dynamically quantifies semantic saliency by integrating feature complexity and motion intensity information from patches. This strategy can adaptively adjust node connectivity based on the computed saliency, thereby enabling the graph structure to focus on information-dense regions such as hands and faces. For temporal dependency modeling, we abandon the conventional pairwise frame interactions and propose a temporal hypergraph construction method. This method employs a learnable clustering algorithm to aggregate semantically correlated nodes within temporal windows into hyperedges, thereby explicitly capturing high-order associations within individual gesture actions that span multiple frames. Extensive experiments on the PHOENIX14, PHOENIX14-T, and CSL-Daily datasets demonstrate that HyperSign outperforms the state-of-the-art (SOTA) approaches in CSLR without any additional annotation information, establishing a new feature learning paradigm for the CSLR task.

To accelerate magnetic resonance imaging (MRI), image reconstruction from under-sampled measurements has been widely used. Recently, the convolutional-Transformer hybrid architecture has dominated the field of MRI reconstruction. To improve calculation performance, two multi-scale (MS) strategies are usually adopted: the one imposed in the intra-cascades in a U-shape style and the one lying in the inter-cascades in a pyramid manner. The two MS strategies have their own benefits but have not been combined together for boosting performance. In this work, we proposed a dynamic Hybrid Unrolled Multi-Scale Network (dHUMUS-Net) by incorporating the two MS strategies together. A novel Optimal Scale Prediction Network is presented to dynamically estimate the optimal scales for all cascades of dHUMUS-Net. Experiments on the fastMRI dataset demonstrate the effectiveness of our method over the state-of-the-art methods.

Recent studies on multi-contrast MRI reconstruction have demonstrated the potential of further accelerating MRI acquisition by exploiting correlation between contrasts. Most of the state-of-the-art approaches have achieved improvement through the development of network architectures for fixed under-sampling patterns, without considering inter-contrast correlation in the under-sampling pattern design. On the other hand, sampling pattern learning methods have shown better reconstruction performance than those with fixed under-sampling patterns. However, most under-sampling pattern learning algorithms are designed for single contrast MRI without exploiting complementary information between contrasts. To this end, we propose a framework to optimize the under-sampling pattern of a target MRI contrast which complements the acquired fully-sampled reference contrast. Specifically, a novel image synthesis network is introduced to extract the redundant information contained in the reference contrast, which is exploited in the subsequent joint pattern optimization and reconstruction network. We have demonstrated superior performance of our learned under-sampling patterns on both public and in-house datasets, compared to the commonly used under-sampling patterns and state-of-the-art methods that jointly optimize the reconstruction network and the under-sampling patterns, up to 8-fold under-sampling factor.

To enhance the nonlinearity of neural networks and increase their mapping abilities between the inputs and response variables, activation functions play a crucial role to model more complex relationships and patterns in the data. In this work, a novel methodology is proposed to adaptively customize activation functions only by adding very few parameters to the traditional activation functions such as Sigmoid, Tanh, and rectified linear unit (ReLU). To verify the effectiveness of the proposed methodology, some theoretical and experimental analysis on accelerating the convergence and improving the performance is presented, and a series of experiments are conducted based on various network models (such as AlexNet, VggNet, GoogLeNet, ResNet and DenseNet), and various datasets (such as CIFAR10, CIFAR100, miniImageNet, PASCAL VOC, and COCO). To further verify the validity and suitability in various optimization strategies and usage scenarios, some comparison experiments are also implemented among different optimization strategies (such as SGD, Momentum, AdaGrad, AdaDelta, and ADAM) and different recognition tasks such as classification and detection. The results show that the proposed methodology is very simple but with significant performance in convergence speed, precision, and generalization, and it can surpass other popular methods such as ReLU and adaptive functions such as Swish in almost all experiments in terms of overall performance.

A major challenge in the analysis of tissue imaging data is cell segmentation, the task of identifying precisely the boundary of each cell in a microscopic image. The cell segmentation task is still challenging due to the variable shapes, large size differences, uneven grayscale, and dense distribution of biological cells in microscopic images. In this paper, we propose a joint feature learning method that integrates the density and boundary branch into a multi-scale convolutional U-Net (MC-Unet). To enhance the supervision of cell density and boundary detection, the density and boundary loss is constructed to guide the joint learning of multiple features, where the density loss branch can address the challenges posed by high density, while the boundary loss branch can address the problems of unclear cell boundaries and partial cell occlusion. A series of experiments on different cell datasets show that two auxiliary branches improve the learning of features on cell density and cell boundaries and that the proposed method is effective on different segmentation models. The code is available at: https://github.com/HuHaigen/Joint-Feature-Learning-for-Cell-Segmentation.

Deep convolutional neural networks (CNNs) have witnessed a great progress in visual recognition over the past years. However, deep CNN models are still suffering from the domain adaptation problem. Most of the existing methods try to resolve this issue by creating more useful samples in the source domain for network training so that the well-trained CNN models can well adapt to more possible variations in the target domain. However, such methods are different with human visual mechanism. Human eyes can effectively recognize images with large variations that were never seen before, as long as human eyes are very familiar with partial contents of the input images. We simulate the visual mechanism of human eyes and make feature responses diverse as far as possible. We proposed a novel angular diversity loss, which contains a pair of angular Spatial Activation Diversity (A-SAD) losses by borrowing the idea of the angular losses. Besides concerning the recognition accuracy, we also focus on understanding deep CNNs. Recent works further pushed the interpretability into the training stage of the CNN models. This helps CNN models learn more meaningful feature representations. Extensive experiment on MNIST dataset and its six variation dataset show the effectiveness of the proposed A-SAD loss.

An activation function is a key component of artificial neural networks (ANNs). It has a great impact on the performance and convergence of neural networks. In this work, a self-adapting methodology is proposed to explore the optimal adaptive activation functions for various tasks based on S-shaped or ReLu-shaped activation functions, which are regulated only by introducing several parameters. To verify the effectiveness of the proposed methodology, a series of comparison experiments are performed with MLP, CNN and RNN network structure on the benchmark datasets of image, text and audio. The experimental results are encouraging, and show that the proposed methodology can locate the optimal activation functions for various tasks. Nevertheless, the obtained functions are competitive and the improvements on network performance are significant compared with other popular activation functions, such as ELU, PReLU, ReLU, and Sigmoid.

For hyperspectral image (HSI) classification, it is challenging to adopt the methodology of sparse-representation-based classification. In this paper, we first propose an l1-minimization-based spectral-spatial classification method for HSIs via a spatial translation-invariant wavelet (STIW)-based sparse representation (STIW-SR), wherein both the spectrum dictionary and the analyzed signal are formed with STIW features. Due to the capability of a STIW to reduce both the observation noise and the spatial nonstationarity while maintaining the ideal spectra, which is proved with our signal-interference-noise spectrum model involved, it is expected that the pixels in the same class congregate in a lower dimensional subspace, and the separations among class-specific subspaces are enhanced, thus yielding a highly discriminative sparse representation. Then, we develop an approach to evaluate the sparsity recoverability of an l1-minimization on HSIs in a probabilistic framework. This approach takes into account not only the recovery probability under the given support length of the l0-norm solution but also the apriori probability of the support length; consequently, it overcomes the inability of traditional mutual/cumulative coherence conditions to address high-coherence HSIs. This paper reveals that the higher sparsity recoverability of a STIW-SR leads to its higher classification accuracy and that the increasing coherence does not necessarily lead to a reduced sparsity recovery probability, and this paper verifies the connection between l0and l1-minimizations on HSIs. Experimental results from realworld HSIs suggest that our classification method significantly outperforms several representative spectral-spatial classifiers and support vector machines.

Face recognition with illumination or pose variation is a challenging problem in image processing and pattern recognition. A novel algorithm using band-reweighed Gabor kernel embedding to deal with the problem is proposed in this paper. For a given image, it is first transformed by a group of Gabor filters, which output Gabor features using different orientation and scale parameters. Fisher scoring function is used to measure the importance of features in each band, and then, the features with the largest scores are preserved for saving memory requirements. The reduced bands are combined by a vector, which is determined by a weighted kernel discriminant criterion and solved by a constrained quadratic programming method, and then, the weighted sum of these nonlinear bands is defined as the similarity between two images. Compared with existing concatenation-based Gabor feature representation and the uniformly weighted similarity calculation approaches, our method provides a new way to use Gabor features for face recognition and presents a reasonable interpretation for highlighting discriminant orientations and scales. The minimum Mahalanobis distance considering the spatial correlations within the data is exploited for feature matching, and the graphical lasso is used therein for directly estimating the sparse inverse covariance matrix. Experiments using benchmark databases show that our new algorithm improves the recognition results and obtains competitive performance.

Face recognition with occlusion is common in the real world. Inspired by the works of structured sparse representation, we try to explore the structure of the error incurred by occlusion from two aspects: the error morphology and the error distribution. Since human beings recognize the occlusion mainly according to its region shape or profile without knowing accurately what the occlusion is, we argue that the shape of the occlusion is also an important feature. We propose a morphological graph model to describe the morphological structure of the error. Due to the uncertainty of the occlusion, the distribution of the error incurred by occlusion is also uncertain. However, we observe that the unoccluded part and the occluded part of the error measured by the correntropy induced metric follow the exponential distribution, respectively. Incorporating the two aspects of the error structure, we propose the structured sparse error coding for face recognition with occlusion. Our extensive experiments demonstrate that the proposed method is more stable and has higher breakdown point in dealing with the occlusion problems in face recognition as compared to the related state-of-the-art methods, especially for the extreme situation, such as the high level occlusion and the low feature dimension.

Detecting protein complexes from protein interaction networks is one major task in the postgenome era. Previous developed computational algorithms identifying complexes mainly focus on graph partition or dense region finding. Most of these traditional algorithms cannot discover overlapping complexes which really exist in the protein-protein interaction (PPI) networks. Even if some density-based methods have been developed to identify overlapping complexes, they are not able to discover complexes that include peripheral proteins. In this study, motivated by recent successful application of generative network model to describe the generation process of PPI networks and to detect communities from social networks, we develop a regularized sparse generative network model (RSGNM), by adding another process that generates propensities using exponential distribution and incorporating Laplacian regularizer into an existing generative network model, for protein complexes identification. By assuming that the propensities are generated using exponential distribution, the estimators of propensities will be sparse, which not only has good biological interpretation but also helps to control the overlapping rate among detected complexes. And the Laplacian regularizer will lead to the estimators of propensities more smooth on interaction networks. Experimental results on three yeast PPI networks show that RSGNM outperforms six previous competing algorithms in terms of the quality of detected complexes. In addition, RSGNM is able to detect overlapping complexes and complexes including peripheral proteins simultaneously. These results give new insights about the importance of generative network models in protein complexes identification.