Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文

Judgement and Disposal of Academic Misconduct Article Copyright Transfer Agreement Proof of Confidentiality Requirements for Electronic Version More.... Chinese Association of Automation National ResearchCenter for Intelligent Computing System Institute of Intelligent Machines,Chinese Academy of Sciences More....

2023 Vol.36 Issue.9, Published 2023-09-25 Papers and Reports    Deep Learning Based Pattern Classification and Detection        Deep Learning Based Pattern Classification and Detection 767 Infrared Small Target Detection Network Inspired by High-Order Differential Equation ZHANG Mingjin, ZANG Fan, YUE Ke, XU Jiamin, LI Yunsong, GAO Xinbo In the fields of infrared detection and infrared tracking, infrared small target detection is widely applied. However, infrared small target detection poses significant challenges. The existing methods for infrared small target detection fail to address complex background issues while losing detailed information during feature extraction. Therefore, an infrared small target detection network inspired by high-order differential equations is proposed. Under the guidance of the interpretable theory, a fourth-order Adams-guided feature fusion module is designed, incorporating adaptive weight factors to effectively fuse multi-scale information from different levels. High-order difference equations are employed to eliminate redundant information through deep learning. The target feature enhancement module utilizes a residual structure composed of convolutions at different scales to suppress background noise and enhance multi-scale features with high information content. Experiments for small target detection on publicly available SIRST dataset show that the proposed network has advantages in the evaluation metrics and visual quality. 2023 Vol. 36 (9): 767-777 [Abstract] ( 344 ) [HTML 1KB] [ PDF 1792KB] ( 444 ) 778 Fuzzy Logic Guided Deep Neural Network with Multi-granularity ZHOU Tianyi, DING Weiping, HUANG Jiashuang, JU Hengrong, JIANG Shu, WANG Haipeng The accurate identification and classification of pathological images are crucial for early disease detection and treatment. In the process of diagnosis, pathologists typically employ a multi-level approach, observing abnormal cell regions at various magnifications. However, existing models often extract features at a single granularity, neglecting the multi-granularity nature of cells. Therefore, a fuzzy logic guided deep neural network with multi-granularity is proposed in this paper. Firstly, multi-granularity feature extraction is conducted for cell structures at three levels of granularity-coarse, medium and fine, and thus the information in pathological tissue images is fully utilized. Additionally, to address the issue of key information redundancy during multi-granularity feature extraction, fuzzy logic theory is introduced. Multiple fuzzy membership functions are set to describe cell features from different angles. Subsequently, fuzzy operations are employed to fuse into a fuzzy universal feature. A fuzzy logic guided cross attention mechanism is designed to guide the multi-granularity features by the universal fuzzy feature. Finally, an encoder is utilized to propagate the features to all patch tokens, resulting in improved classification accuracy and robustness. Experiments demonstrate that the proposed network achieves high accuracy in the classification of pathological images. 2023 Vol. 36 (9): 778-792 [Abstract] ( 215 ) [HTML 1KB] [ PDF 2071KB] ( 520 ) 793 Object Detection Algorithm with Dual-Modal Rectification Fusion Based on Self-Guided Attention ZHANG Jinglei, GONG Wenhao, JIA Xin The traditional dual-modal object detection algorithms struggle to overcome low-contrast noise in complex scenes, such as fog, glare and dark night, and they cannot recognize small-size objects effectively. To solve these problems, an object detection algorithm with dual-modal rectification fusion based on self-guided attention is proposed. Firstly, a dual-modal fusion network is designed to rectify the low-contrast noise in the input images(visible and infrared images) by channel and spatial feature rectification. Consequently, the complementary information is acquired from the rectified features to accurately achieve feature fusion and the detection accuracy of the algorithm in the complex scenes is improved. Secondly, a self-guided attention mechanism is established to learn the dependency among pixels in the images. Thus, the fusion capability of features at different scales and the detection accuracy of the algorithm for small-scale objects are improved. Extensive experiments on six datasets, including pedestrian datasets, pedestrian-vehicle datasets and aerial vehicle datasets, demonstrate the superiority of the proposed approach. 2023 Vol. 36 (9): 793-805 [Abstract] ( 254 ) [HTML 1KB] [ PDF 3392KB] ( 492 ) 806 Temporal Action Unit Perception Based Open Set Action Recognition YANG Kaixiang, GAO Junyu, FENG Yangbo, XU Changsheng In open set action recognition tasks, a model is requested to identify categories within the training set accurately and reject unknown actions that never appear in the training set. Currently, most of the methods treat the action as a whole, ignoring the fact that the action can be decomposed into finer-grained action units. To address this issue, a method for temporal action unit perception based open set action recognition is proposed in this paper. Firstly, an action unit relationship module is designed to learn fine-grained features of action units, and thus the relational pattern between actions and action units is obtained. The unknown actions are identified according to the different degrees of activation of known and unknown actions on action units. Secondly, an action unit temporal module is designed to model the temporal information of action units. The temporal characteristics of action units are explored to further distinguish between known actions and unknown actions that are visually similar but confusable with each other. Finally, with comprehensive consideration of both relational patterns and temporal information of action units, the model is equipped with the capability of distinguishing known actions from unknown actions. Experimental results on three action recognition datasets demonstrate the superior performance of the proposed method. 2023 Vol. 36 (9): 806-817 [Abstract] ( 179 ) [HTML 1KB] [ PDF 1350KB] ( 357 ) 818 Remote Sensing Image Recognition Algorithm Based on Pseudo Global Swin Transformer WANG Keping, ZUO Xinhao, YANG Yi, FEI Shumin Determining the core target aligning with human thinking habits in the context of multiple concurrent targets is one of the key factors in remote sensing image recognition. Therefore,the effective allocation of attention in accordance with human visual habits in a global perspective is one of the ways to select core targets. In this paper, combining the concept of extracting features using the Transformer and the advantages of the Swin Transformer in reducing computational complexity through image gridding, a remote sensing image recognition algorithm based on pseudo global Swin Transformer is proposed.The pseudo global Swin Transformer module is built to aggregate the local information of rasterized remote sensing images into a single feature value, replacing the pixel-based global information to obtain global features with smaller computational cost, and thus the perceptual ability of the model for all targets is effectively improved. Meanwhile, by introducing a receptive field adaptive scaling module based on deformable convolutions, the receptive field is shifted towards core targets to enhance the network attention to core target information and then achieve precise recognition of remote sensing images. Experiments on RSSCN7, AID, and OPTIMAL-31 remote sensing image datasets show that the proposed algorithm achieves high recognition accuracy and parameter identification efficiency. 2023 Vol. 36 (9): 818-831 [Abstract] ( 218 ) [HTML 1KB] [ PDF 5148KB] ( 298 ) 832 Nonparametric Image Clustering Based on Variational Bayesian Contrastive Network ZHANG Shengjie, WANG Yifei, XIANG Wang, XUE Dizhan, QIAN Shengsheng The number of clusters in nonparametric image clustering is unknown and it needs to be discovered by the model automatically. Although some existing Bayesian methods can automatically infer the number of clusters, they are not feasible on large-scale image datasets due to the high computational costs or over-reliance on learned features. Therefore, nonparametric image clustering based on variational Bayesian contrastive network is proposed in this paper. Firstly, image features are extracted by ResNet. Secondly, deep variational Dirichlet process mixture is put forward to automatically infer the number of clusters, and it can be directly embedded into end-to-end deep models and jointly optimized with feature extractors. Finally, polarized contrast clustering learning is presented, and the denoising strategy with polarized label is utilized to denoise and polarize the labels. The polarized labels and data augmented predicted labels are employed for comparative learning to jointly optimize image feature extractors and clustering model. Experiments on three benchmark datasets show that the performance of the proposed method is superior. 2023 Vol. 36 (9): 832-841 [Abstract] ( 167 ) [HTML 1KB] [ PDF 2016KB] ( 420 ) Papers and Reports 842 Hyperbolic Positive Definite Kernels Based on Möbius Gyrovector Space YANG Meimei, FANG Pengfei, ZHU Shipeng, XUE Hui Hierarchical data is widely present in various machine learning scenarios and the data can be encoded in hyperbolic spaces with very low distortion. Kernel methods are introduced to further enhance the representation capability of hyperbolic space. However, the existing hyperbolic kernels still have the drawbacks of low adaptive capacity or data distortion. To address these issues, hyperbolic positive definite kernels based on Möbius gyrovector space is proposed in this paper. By leveraging the relationship between the Möbius gyrovector space and the Poincaré model, a class of hyperbolic kernel functions, the Möbius radial basis kernels, are constructed. Specifically, the Möbius gyrodistance is employed in place of the Euclidean distance to construct the Möbius Gaussian kernel and the Möbius Laplacian kernel, with the positive definiteness of the kernel functions further demonstrated. Moreover, kernel functions are transformed from complex space to real space, and thus they are more suitable for most machine learning tasks. Experiments on several real-world social network datasets validate the effectiveness of the proposed method. 2023 Vol. 36 (9): 842-855 [Abstract] ( 145 ) [HTML 1KB] [ PDF 822KB] ( 172 ) 856 Regularization Optimization Algorithm for Heterogeneous Data FederatedLearning Model Based on Structure Enhancement ZHANG Min, LIANG Meiyu, XUE Zhe, GUAN Zeli, PAN Zhenhui, ZHAO Zehua In federated learning, due to the heterogeneous distribution of local data among different clients, the optimization objectives of client models trained on local datasets are inconsistent with the global model, leading to client drift and affecting the performance of global model. To address the issue of performance decline or even divergence in federated learning models caused by non-independently and identically distributed data, a regularization optimization algorithm for heterogeneous data federated learning model based on structure enhancement(FedSER) is proposed from the perspective of the generality of local models. While training on local data with heterogeneous distributions, clients sample subnetworks in a structured manner. Local data of client are augmented, and different subnetworks are trained with the augmented data to learn enhanced representations, resulting in more generalized client network models. The models counteract the client drift caused by the heterogeneity of local data and achieve a better global model in federated aggregation. Extensive experiments on the CIFAR-10, CIFAR-100 and ImageNet-200 datasets demonstrate the superior performance of FedSER. 2023 Vol. 36 (9): 856-865 [Abstract] ( 150 ) [HTML 1KB] [ PDF 967KB] ( 325 )

Supervised by China Association for Science and Technology Sponsored by Chinese Association of Automation NationalResearchCenter for Intelligent Computing System Institute of Intelligent Machines, Chinese Academy of Sciences Published by Science Press

Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn Supported by Beijing Magtech 　Email:support@magtech.com.cn