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....

2025 Vol.38 Issue.10, Published 2025-10-25 Papers and Reports    Researches and Applications        Papers and Reports 861 Negative-Free and Weight-Decorrelated Contrastive Network for Graph-Agnostic Clustering BAI Shengxing, ZHANG Yuhong, ZHOU Peng, WU Xindong Existing contrastive deep graph clustering methods heavily rely on the homophily assumption of the input graph, and thereby result in false negatives in negative sampling and feature redundancy in heterophilous graphs. These problems degrade clustering performance. To solve these problems, a graph-agnostic clustering framework, negative-free and weight-decorrelated contrastive network（NFWD）, is proposed. First, a feature graph is constructed as a supplementary view by node attribute similarity. Node representations from the feature graph and the original graph are obtained via Laplacian smoothing filters and a shared-parameter multilayer perceptron, respectively. Consequently, the reliance of the original graph on the homophily assumption is significantly reduced. Second, to tackle the false negative problem caused by class conflicts in heterophilous graphs, cluster information is derived from adaptively fused node representations for the construction of cluster-central node representations. Then, a negative-free strategy combining node-level and cluster-level feature contrast is proposed to effectively mitigate this problem. Finally, an orthogonal constraint is applied to the weight matrix of MLP to actively suppress redundant features. Experiments on six benchmark graph datasets demonstrate the effectiveness and robustness of NFWD in graph-agnostic scenarios. 2025 Vol. 38 (10): 861-875 [Abstract] ( 60 ) [HTML 1KB] [ PDF 1890KB] ( 46 ) 876 Bionic Path Integration Model for Mobile Robots Inspired by Entorhinal-Hippocampal Structure of Rat Brain LIAO Yishen, YU Hejie, YU Naigong, WANG Chenghua, FU Shufei Path integration is recognized as one of the key neural mechanisms underlying spatial navigation in mammals. A bionic path integration model for mobile robots inspired by entorhinal-hippocampal structure of the rat brain（EHPI） is proposed in this paper. EHPI provides an efficient and biologically interpretable solution for autonomous localization of mobile robots in environments without external reference positioning. Taking self-motion cues as input, EHPI fully emulates the hierarchical information processing of spatial cells, including theta cells, grid cells, place cells, and boundary cells. First, continuous dynamic integration signals are generated by coupling the real-time velocity and heading of the robot with hippocampal theta rhythms. Next, multi-layer grid neural sheets are constructed to simulate grid cell populations with different scales and orientations. Connection weights are dynamically adjusted via online competitive Hebbian learning to select and output the grid signals with the highest current phase consistency. Finally, place cells integrate the aforementioned two types of signals to form stable unimodal firing fields, while boundary cells detect the boundaries of the current encoding area to trigger periodic resetting. Thereby, stable positional representation in spaces of arbitrary scale is achieved. Experimental results demonstrate that EHPI achieves superior performance with a small average localization error. 2025 Vol. 38 (10): 876-892 [Abstract] ( 22 ) [HTML 1KB] [ PDF 5925KB] ( 33 ) 893 Wavelet-Enhanced Guidance for Diffusion-Based Sequential Recommendation ZHOU Xi, XIA Hongbin, WANG Xiaofeng Since most conditional diffusion-based sequential recommendation models directly extract guiding signals from the historical interaction sequences of users, the generated signals are susceptible to noise and lack sufficient contextual information, thereby limiting the generation capability of the model. To address these issues, wavelet-enhanced guidance for diffusion-based sequential recommendation（WEG4Rec） is proposed in this paper. First, multi-frequency segmentation results of historical interaction embeddings are obtained via the wavelet transform. On this basis, adaptive dimensional projection and linear attention are introduced to generate multi-granularity interest embeddings. Second, the multi-granularity interest embeddings are employed to guide the reverse reconstruction process of the diffusion model. Finally, a multi-task strategy is adopted to jointly optimize the recommendation model during the training. Extensive experiments on four real-world datasets demonstrate the superior performance of WEG4Rec. 2025 Vol. 38 (10): 893-910 [Abstract] ( 30 ) [HTML 1KB] [ PDF 1248KB] ( 25 ) Researches and Applications 911 Lightweight Unsupervised Multi-exposure Light Field Image Fusion Based on Full-Aperture Estimation LI Yulong, CHEN Yeyao, JIN Chongchong, JIANG Gangyi Multi-exposure light field（LF） image fusion is an effective way to overcome the limited dynamic range of LF cameras. However, due to the high-dimensional structure of LFs, existing methods struggle to efficiently process multi-exposure LF images. To address this issue, a method for lightweight unsupervised multi-exposure LF image fusion based on full-aperture estimation（MELFF-FAE） is proposed. First, the representative scene information is extracted from the central sub-aperture image（SAI） to reduce the heavy computational burden caused by the input of the full LF image. Second, a full-aperture weight estimation module is designed to obtain the fusion weight of the full LF image by mining the LF angular information. The difference between the boundary SAIs and the central SAI is utilized to construct a full weight map in the feature space. Finally, the weight map is multiplied with the source image to generate a fused LF image. Experimental results demonstrate that MELFF-FAE can generate LF images with high contrast and detailed textures while preserving good angular consistency. Moreover, compared to existing representative methods, MELFF-FAE achieves superior results in both quantitative and qualitative comparisons while significantly reducing the computational burden. 2025 Vol. 38 (10): 911-924 [Abstract] ( 26 ) [HTML 1KB] [ PDF 8690KB] ( 37 ) 925 Sequential Recommendation with Large Language Model Enhancement and Collaborative Information Fusion ZHA Longbao, HUANG Qi, WANG Mingwen, ZHOU Junxiang, LUO Wenbing Existing sequential recommendation methods fail to fully explore item attribute semantic information and suffer from a semantic space migration mismatch. These limitations result in inadequate recommendation capabilities of the existing methods for long-tail items. To address this issue, a method for sequential recommendation with large language model enhancement and collaborative information fusion（LLM-CFSR） is proposed in this paper. First, fine-grained semantic embeddings are generated with a large language model through attribute-level data augmentation and contrastive fine-tuning techniques to capture the deep semantic associations of long-tail items. Then, a dual-view modeling framework is designed to jointly model user preferences from both semantic and collaborative views. Finally, to promote deep interaction between semantic information and collaborative signals, a cross-attention mechanism is introduced to achieve multi-level information fusion across embedding layers, sequence layers, and prediction layers. Experimental results on Yelp, Amazon Fashion and Amazon Beauty datasets demonstrate that LLM-CFSR improves the overall recommendation performance and the long-tail item recommendation performance. 2025 Vol. 38 (10): 925-937 [Abstract] ( 33 ) [HTML 1KB] [ PDF 906KB] ( 25 ) 938 Clean-Label Backdoor Watermarking for Face Recognition Models YAN Xing, LI Yun Face recognition models are widely applied in critical areas, such as security authentication and intelligent surveillance. These models are faced with significant security and copyright risks due to their high reliance on sensitive biometric features. Backdoor watermarking technology for face recognition models is widely utilized for copyright verification, but most existing methods rely on dirty-label strategies. Consequently, data semantic consistency is destroyed, and the watermarks can be easily detected by current backdoor-detection mechanisms, which limit practical deployment. To address these issues, a clean-label backdoor watermarking method for face recognition models（CBW2F） is proposed in this paper. High imperceptibility and strong robustness are achieved without modifying any sample labels. Specifically, imperceptible adversarial perturbations are first applied to a subset of samples. The model dependence on original salient features is weakened, and the learning of the embedded backdoor trigger pattern is encouraged. A structured and visually natural rainbow filter is then introduced as the trigger. Through its cooperation with the perturbation, the model achieves effective watermark embedding while maintaining its original recognition performance. Experiments demonstrate that CBW2F effectively evades label-consistency-based backdoor detection and maintains strong robustness under various watermark removal attacks, including model fine-tuning and model distillation. It outperforms existing state-of-the-art approaches across multiple evaluation metrics, providing a practical solution for copyright protection in face recognition models. 2025 Vol. 38 (10): 938-948 [Abstract] ( 30 ) [HTML 1KB] [ PDF 1859KB] ( 28 ) 949 Chinese-Burmese Neural Machine Translation Model Based on Attention-Optimized Adversarial Training LAI Hua, LI Yanduo, ZHANG Siqi, LI Ying, YU Zhengtao, MAO Cunli, HUANG Yuxin Excessive noise introduction during adversarial training can degrade the robustness of translation models. To address this issue, a Chinese-Burmese neural machine translation method based on attention-optimized adversarial training is proposed. In the training phase, white-box adversarial attacks are utilized to generate perturbation samples along the gradient direction. A mixed attention weight filtering strategy is introduced to prioritize perturbations on words that produce a greater impact on translation quality, thereby improving the specificity of perturbations without increasing the overall noise ratio. During the inference phase, a relative entropy loss is employed to effectively narrow the gap between noisy and clean distributions and balance the robustness of the model to noise and its fitting ability on clean data. Experiments on the Burmese-Chinese translation task demonstrate that the proposed method achieves a significant improvement over multiple baseline models. 2025 Vol. 38 (10): 949-959 [Abstract] ( 28 ) [HTML 1KB] [ PDF 1118KB] ( 39 )

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