Search

Hypergraph neural networks (HGNNs), extending the techniques of graph neural networks, have been applied to various fields due to their ability to capture more complex high-order node relationships. However, for hyperspectral image (HSI) classification tasks, previous HGNN-based works usually constructed hypergraphs using pixels as nodes, resulting in massive computational costs. Meanwhile, pixel-level personalized features are required for HSI classification. To achieve high efficiency and accuracy simultaneously, this paper presents a fast hypergraph neural network with detail preservation (DPFHNet) for HSI classification. It constructs hypergraphs at the superpixel level to reduce time consumption and supplement pixel-level detail features through a classification refinement module. This framework contains multiple stages. Firstly, its main stage is designed with HGNNs from a superpixel viewpoint rather than pixels, providing a fast strategy to capture high-order complex relationships of multiple homogeneous irregular regions. After that, auxiliary stages based on convolutional neural networks are integrated into the main stage, which adopts a hierarchical design and attempts to acquire pixel-level spatial-spectral information before the hypergraph feature extraction of the main stage, assisting in learning more valuable features. Finally, a classification refinement module is constructed, which generates pixel-level detail features to refine the superpixel-level features obtained by HGNN. Experiments on three datasets illustrate that DPFHNet achieves competitive results and efficiency compared to advanced methodologies. [ABSTRACT FROM AUTHOR]

Vision Transformer (ViT) has attracted tremendous attention and achieved remarkable success on high-level visual tasks. However, ViT relies on costly pre-training on large external datasets and is strict in data and calculations, making it an obstacle to running on common equipment. To address this challenge, we propose a simple and efficient Transformer namely SRT tailored for the image super-resolution (SR) reconstruction task. It is trained on a single GPU card without large-scale pre-training. At the beginning of the whole model, we introduce a convolutional stem module instead of straightforward tokenization of raw input images for low-level feature extraction and steady training. In the main Transformer learning phase, we exploit an additional head-convolution to make up for the lack of information interaction in multi-head self-attention (MHSA). Then further to strengthen the spatial correlation of neighboring tokens in MLP, a locally-enhanced feed-forward layer is thus employed to promote local dependencies. In terms of the inefficiency of Transformer, a channel reduction strategy is presented in MHSA, which dramatically reduces the complexity and space resources. Experimental results demonstrate the proposed Transformer model can rival the current state-of-the-art methods with a single GPU card. [ABSTRACT FROM AUTHOR]

Several studies have shown the excellent performance of deep learning in image segmentation. Usually, this benefits from a large amount of annotated data. Medical image segmentation is challenging, however, since there is always a scarcity of annotated data. This study constructs a novel deep network for medical image segmentation, referred to as asymmetric U-Net generative adversarial networks with multi-discriminators (AU-MultiGAN). Specifically, the asymmetric U-Net is designed to produce multiple segmentation maps simultaneously and use the dual-dilated blocks in the feature extraction stage only. Further, the multi-discriminator module is embedded into the asymmetric U-Net structure, which can capture the available information of samples sufficiently and thereby promote the information transmission of features. A hybrid loss by the combination of segmentation and discriminator losses is developed, and an adaptive method of selecting the scale factors is devised for this new loss. More importantly, the convergence of the proposed model is proved mathematically. The proposed AU-MultiGAN approach is implemented on some standard medical image benchmarks. Experimental results show that the proposed architecture can be successfully applied to medical image segmentation, and obtain superior performance in comparison with the state-of-the-art baselines. [ABSTRACT FROM AUTHOR]

The classification of 3D point clouds is a regular task, but remains a highly challenging problem because 3D point clouds usually contain a large amount of information on irregular shapes. Several recent studies have shown the excellent performance of deep learning in 3D point cloud classification. Convolutional neural network (CNN)-based 3D point cloud classification methods are also increasingly used owing to their efficient and convenient feature extraction capability. However, most of these methods do not take much prior information and local structural information into consideration, often resulting in their inability to extract sufficient information to improve the classification accuracy. In this study, we present a novel convolution operation named HyConv, which includes two key components. First, inspired by 2D convolution, we design a feature transformation module to capture more local structural information. Second, to extract the prior information, a hybrid weight module is introduced to estimate two types of weights on the basis of the distribution information of the spatial and feature domains. Additionally, we propose an adaptive method to learn hybrid weights to obtain hybrid distribution information. Finally, based on the proposed convolutional operator HyConv, we build a deep neural network Hybrid-CNN and conduct experiments on two commonly used datasets. The results show that our hybrid network outperforms most existing methods on ModelNet40. Furthermore, state-of-the-art performance is achieved with ScanObjectNN, which is a great improvement compared with existing methods. [ABSTRACT FROM AUTHOR]

Inspired by the perceived saturation of human visual system, this paper proposes a two-stream hybrid networks to simulate binocular vision for salient object detection (SOD). Each stream in our system consists of unsupervised and supervised methods to form a two-branch module, so as to model the interaction between human intuition and memory. The two-branch module parallel processes visual information with bottom-up and top-down SODs, and output two initial saliency maps. Then a polyharmonic neural network with random-weight (PNNRW) is utilized to fuse two-branch’s perception and refine the salient objects by learning online via multi-source cues. Depend on visual perceptual saturation, we can select optimal parameter of superpixel for unsupervised branch, locate sampling regions for PNNRW, and construct a positive feedback loop to facilitate perception saturated after the perception fusion. By comparing the binary outputs of the two-stream, the pixel annotation of predicted object with high saturation degree could be taken as new training samples. The presented method constitutes a semi-supervised learning framework actually. Supervised branches only need to be pre-trained initial, the system can collect the training samples with high confidence level and then train new models by itself. Extensive experiments show that the new framework can improve performance of the existing SOD methods, that exceeds the state-of-the-art methods in six popular benchmarks. [ABSTRACT FROM AUTHOR]

In recent years, convolutional neural networks have been successfully applied to single image super-resolution (SISR) tasks, making breakthrough progress both in accuracy and speed. In this work, an improved dual-scale residual network (IDSRN), achieving promising reconstruction performance without sacrificing too much calculations, is proposed for SISR. The proposed network extracts features through two independent parallel branches: dual-scale feature extraction branch and texture attention branch. The improved dual-scale residual block (IDSRB) combined with active weighted mapping strategy constitutes the dual-scale feature extraction branch, which aims to capture dual-scale features of the image. As regards the texture attention branch, an encoder–decoder network employing symmetric full convolutional-deconvolution structure acts as a feature selector to enhance the high-frequency details. The integration of two branches reaches the goal of capturing dual-scale features with high-frequency information. Comparative experiments and extensive studies indicate that the proposed IDSRN can catch up with the state-of-the-art approaches in terms of accuracy and efficiency. [ABSTRACT FROM AUTHOR]

In this article, a new human face recognition scheme is proposed. The proposed system is based on the sparsity and fuzziness, and utilizes independent component analysis (ICA). The scheme includes four parts: a fuzzy comprehensive judgment model for estimating whether the information carried by training samples is enough or not, a proper edge extraction operator to discover more hidden information for single image, ICA feature extractor, and a sparse representation model for correlation coefficient calculation to classify testing samples. In view of the intrinsic patterns of gray information distribution of face images, a weighted fuzzy distance for judgement model and cluster analysis is proposed. The new proposed method is tested on ORL, FERET and UMIST databases. The experiment results demonstrate and illustrate the feasibility of the proposed method and the effective performances on recognition rate. [ABSTRACT FROM AUTHOR]

Recent studies have shown that super-resolution can be significantly improved by using deep convolution neural network. Although applying a larger number of convolution kernels can extract more features, increasing the number of feature mappings will dramatically increase the training parameters and time complexity. In order to balance the workload among all units and maintain appropriate time complexity, this paper proposes a new network structure for super-resolution. For the sake of making full use of context information, in the structure, the operations of division (S) and fusion (C) are added to the pyramidal bottleneck residual units, and the dense connected methods are used. The proposed network include a preliminary feature extraction net, seven residual units with dense connections, seven convolution layers with the size of 1 × 1 after each residual unit, and a deconvolution layer. The experimental results show that the proposed network has better performance than most existing methods. [ABSTRACT FROM AUTHOR]

Deep learning has shown its great potential in the field of image classification due to its powerful feature extraction ability, which heavily depends on the number of available training samples. However, it is still a huge challenge on how to obtain an effective feature representation and further learn a promising classifier by deep networks when faced with few-shot classification tasks. This paper proposes a multi-features adaptive aggregation meta-learning method with an information enhancer for few-shot classification tasks, referred to as MFAML. It contains three main modules, including a feature extraction module, an information enhancer, and a multi-features adaptive aggregation classifier (MFAAC). During the meta-training stage, the information enhancer comprised of some deconvolutional layers is designed to promote the effective utilization of samples and thereby capturing more valuable information in the process of feature extraction. Simultaneously, the MFAAC module integrates the features from several convolutional layers of the feature extraction module. The obtained features then feed into the similarity module so that implementing the adaptive adjustment of the predicted label. The information enhancer and MFAAC are connected by a hybrid loss, providing an excellent feature representation. During the meta-test stage, the information enhancer is removed and we keep the remaining architecture for fast adaption on the final target task. The whole MFAML framework is solved by the optimization strategy of model-agnostic meta-learner (MAML) and can effectively improve generalization performance. Experimental results on several benchmark datasets demonstrate the superiority of the proposed method over other representative few-shot classification methods. • A multi-features adaptive aggregation meta-learning method is proposed. • An information enhancer is built to capture more valuable information. • A multi-features aggregation classifier is designed in an adaptive way. • A mixed loss is presented by combining the classification and reconstruction losses. [ABSTRACT FROM AUTHOR]

As a class of important classifiers, feedforward neural networks (FNNs) have been used considerably in the study of pattern recognition. Since the inputs to FNNs are usually vectors, and many data are usually presented in the form of matrices, the matrices have to be decomposed into vectors before FNNs are employed. A drawback to this approach is that important information regarding correlations of elements within the original matrices are lost. Unlike traditional vector input based FNNs, a new algorithm of extended FNN with matrix inputs, called two-dimensional back-propagation (2D-BP), is proposed in this paper to classify matrix data directly, which utilizes the technique of incremental gradient descent to fully train the extended FNNs. These kinds of FNNs help to maintain the matrix structure of the 2D input features, which helps with image recognition. Promising experimental results of handwritten digits and face-image classification are provided to demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Point cloud data obtained by 3D scanning equipment is often incomplete. In recent years, to complete the missing point clouds has become an important task in computer visualization research. This paper adopts a coarse-to-fine completion strategy and build a novel adaptive region shape fused network (ARSF-Net), which contains three core modules, namely region shape encoding module (RSE), adaptive feature selection-aggregation module (ASA), and encoding-attention transformer module (EAT). The RSE module adaptively aggregates the latent shape information contained in local features according to the feature strength. In ASA module, we first treat point coordinates and shape features as parent nodes and design a hybrid correlation method to adaptively group parent nodes. Then, each set of parent nodes generates a child node. Finally, we splice the features and points in the parent node and child node separately to double the number. For EAT module, we learn features from the encoding stage and use a coordinate-based embedding transformer to generate uniform high-resolution point clouds. Compared with previous methods, we pay special attention to the difference among the latent shape information contained in the local point clouds, thus making the local feature extraction more interpretable. At the same time, to generate valid detail features from the original ones, we abundantly consider the correlation among the original ones, and directly combine the original features with the generated ones. Our experiments on different datasets verify the good performance of ARSF-Net in the point cloud completion task. [ABSTRACT FROM AUTHOR]

This paper designs a novel human face recognition method, which is mainly based on a new feature extraction method and an efficient classifier – random weight network (RWN). Its innovation of the feature extraction is embodied in the good fusion of the geometric features and algebraic features of the original image. Here the geometric features are acquired by means of fast discrete curvelet transform (FDCT) and 2-dimensional principal component analysis (2DPCA), while the algebraic features are extracted by a proposed quasi-singular value decomposition (Q-SVD) method that can embody the relations of each image under a unified framework. Subsequently, the efficient RWN is applied to classify these fused features to further improve the recognition rate and the recognition speed. Some comparison experiments are carried out on six famous face databases between our proposed method and some other state-of-the-art methods. The experimental results show that the proposed method has an outstanding superiority in the aspects of separability, recognition rate and training time. [ABSTRACT FROM AUTHOR]