Sunday, May. 11, 2025 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.3, Published 2025-03-25 Papers and Reports    Researches and Applications    Surveys and Reviews        Papers and Reports 193 Role-Based Adaptive Parameter Sharing Method FANG Baofu, WANG Qiong, WANG Hao, WANG Zaijun In large-scale heterogeneous multi-agent reinforcement learning, parameter sharing is often utilized to reduce the number of training parameters and accelerate the training process. However, the traditional full parameter sharing approach is prone to causing excessive behavioral uniformity among agents, while independent parameter training methods are constrained by computational complexity and memory limitations. Therefore, a role-based adaptive parameter sharing(RAPS) method is proposed in this paper. First, agents are grouped into roles based on their task characteristics. Then, within a unified network structure, sparse sub-network structures are generated for different agent roles by integrating unstructured network pruning techniques. A dynamic adjustment mechanism is introduced to adaptively optimize the ratio of shared and independent parameters according to task requirements. Additionally, a collaborative loss function between roles is incorporated to further enhance coordination among heterogeneous agents. Thus, computational complexity is effectively reduced by RAPS while behavioral diversity among heterogeneous agents is preserved. Experimental results demonstrate that RAPS improves the performance and scalability of multi-agent systems significantly in different multi-agent tasks. 2025 Vol. 38 (3): 193-204 [Abstract] ( 63 ) [HTML 1KB] [ PDF 2103KB] ( 48 ) 205 Few-Shot Deepfake Face Detection Method Based on Vision-Language Model YANG Hongyu, LI Xinghang, CHENG Xiang, HU Ze Aiming at the limitations of existing deepfake face detection methods in terms of model complexity, sample size requirements and adaptability to new deepfake techniques, a few-shot deepfake face detection method based on visual-language model(FDFD-VLM) is proposed. FDFD-VLM is built upon contrastive language-image pre-training(CLIP). Visual features are optimized through a face region extraction and high-frequency feature enhancement module. Prompt adaptability is improved by a classless differentiated prompt optimization module, while multimodal feature representation is strengthened by CLIP encoding attention optimization module. Additionally, a triplet loss function is introduced to improve the model discriminative capability. Experimental results demonstrate that FDFD-VLM outperforms existing methods on multiple deepfake face datasets and achieves efficient detection performance in few-shot deepfake face detection scenarios. 2025 Vol. 38 (3): 205-220 [Abstract] ( 48 ) [HTML 1KB] [ PDF 2818KB] ( 35 ) 221 Rough Set Model Based on Fuzzy Purity Granular Ball WANG Xin, HUANG Bing As a classical attribute reduction method, granular ball neighborhood rough set(GBNRS) is constrained by the strict requirement that the purity of granular balls must be exactly 1. As a result, a large number of granular balls with a sample size of 1 are generated at the class boundaries. These granular balls are often misjudged as outliers and eliminated, and the loss of boundary information is caused. To address this issue, a fuzzy purity function is first defined. The function integrates membership degree and class labels as an evaluation metric for the quality of granular balls. Based on dynamic quality assessment and optimization strategies, the function takes into account three aspects: the membership degree of data points, the class labels of data points, and the class labels of granular balls. Nextly, during the granular ball splitting process, a classification significance threshold β is introduced, the m value of M-means is adaptively adjusted, and a granular ball generation method based on fuzzy purity is constructed. Furthermore, for the attribute reduction problem in rough set theory, a forward attribute reduction algorithm is designed, and a rough set model based on fuzzy purity granular ball(FPGBRS) is established. Finally, experiments on 12 real datasets demonstrate that FPGBRS can improve classification accuracy and efficiency. 2025 Vol. 38 (3): 221-232 [Abstract] ( 38 ) [HTML 1KB] [ PDF 807KB] ( 37 ) Surveys and Reviews 233 A Survey of Deep Graph Clustering Methods Based on Different Learning Paradigms ZHOU Lijuan, WU Mengqi, LI Xinran, NIU Changyong Graph clustering aims to partition graph nodes into different categories in an unsupervised manner, facilitating the discovery of hidden patterns, community structures and organizational relationships within complex systems. Existing methods construct different self-supervised information through various learning paradigms to guide graph representation learning and promote clustering. Therefore, the learning paradigm is the key to clustering algorithms. However, few existing reviews discuss graph clustering from the perspective of different learning paradigms. In this paper, the research progress on graph clustering based on different learning paradigms is summarized. Clustering methods are classified into reconstructive graph clustering, contrastive graph clustering, adversarial graph clustering and hybrid graph clustering. Considering the research scope and clustering effect, reconstructive graph clustering and contrastive graph clustering are discussed in detail. Graph clustering results on single-relation and multi-relation datasets are compared. The results show that contrastive graph clustering performs better on single-relation datasets, while reconstructive graph clustering is more effective on multi-relation datasets. Finally, the challenges faced in the graph clustering field are summarized, and future research directions are pointed out as well. The applications of deep graph clustering across various domains are additionally introduced. 2025 Vol. 38 (3): 233-251 [Abstract] ( 41 ) [HTML 1KB] [ PDF 1023KB] ( 36 ) Researches and Applications 252 Parallel Chefs via Agentic Intelligence:From AI Agents to Smart Digital Robotic Cuisine Systems LI Bai, SONG Zihan, LI Xinyuan, HUANG Jun, TIAN Yonglin, YIN Zhuyan, WANG Fei-Yue Progress is made by conversation-based AI due to the rapid development of large language models(LLMs). However, limitations still remain in performing more complex tasks and decision-making. Therefore, agentic intelligence is increasingly emphasized to address the bottleneck that LLMs are confined to information processing. In this paper, a parallel chef cooking system based on agentic intelligence is presented to offer an end-to-end intelligent approach from dish planning to cooking execution. User health data, medical history, and dietary preferences are incorporated to enable personalized recipe design and cooking control. In addition, multi-agent structures are built upon the DeepSeek framework, and specialized Q&A pairs are extracted from culinary literature to offline fine-tune the large language model, thereby imparting cooking reasoning capabilities. Simulation experiments show that compared to GPT o1 pro, a static large model with strong reasoning ability, the agentic approach integrates more extensive professional knowledge and better meets user requirements, showcasing its potential in dietary health and personalized cooking services. 2025 Vol. 38 (3): 252-267 [Abstract] ( 41 ) [HTML 1KB] [ PDF 1373KB] ( 37 ) 268 Zero-Shot Infrared and Visible Image Fusion Based on Fusion Curve LIU Duo, ZHANG Guoyin, SHI Yiqi, TIAN Ye, ZHANG Liguo To solve the problems of color distortion and the loss of thermal target details in infrared and visible image fusion, a method for zero-shot infrared and visible image fusion based on fusion curve(ZSFuCu) is proposed. The fusion task is transformed into an image-specific curve estimation process using a deep network. Texture enhancement and color feature preservation of thermal targets are achieved through pixel-level nonlinear mapping. A multi-dimensional visual perception loss function is designed to construct the constrain mechanism from three perspectives: contrast enhancement, color preservation and spatial continuity. The high-frequency information and color distribution of the fused image are collaboratively optimized with the retention of structural features and key information. The zero-shot training strategy is employed, and the adaptive optimization of parameters can be completed only using a single infrared and visible image pair, which shows strong robustness in fusion across various lighting conditions. Experiments demonstrate that ZSFuCu significantly improves target prominence, detail richness and color naturalness, validating its effectiveness and practicality. 2025 Vol. 38 (3): 268-279 [Abstract] ( 28 ) [HTML 1KB] [ PDF 7367KB] ( 41 ) 280 Steady-State Visual Evoked Potential Detection Algorithm Based on Harmonics Correction and Generalization LÜ Yanhao, LUO Tianjian Steady-state visual evoked potential(SSVEP) is widely utilized in the design of brain-computer interfaces with high information transfer rates(ITR). Existing SSVEP detection algorithms enhance SSVEP components with high signal-noise ratios while suppressing non-SSVEP components by computing optimal spatial filters. However, these algorithms heavily depend on the quality of training samples, resulting in performance degradation in the early stage of SSVEP presentation. To overcome this limitation, a steady-state visual evoked potential detection algorithm based on harmonics correction and generalization(HCG) is proposed. First, the average training templates are corrected by the sine-cosine harmonics reference signals to enhance ITR during the early-middle stage of SSVEP presentation. Subsequently, the task-related component analysis is employed to enhance ITR during the latter stage of SSVEP presentation. For SSVEP detection, the average training templates are weighted and matched by these two types to ensure high ITR during all stages of SSVEP presentation. Comparative experiments are conducted on two public SSVEP datasets. Experiments demonstrate that HCG outperforms the current benchmark algorithms in terms of detection accuracy and ITR, as well as computational efficiency. Moreover, ablation experiments confirm that the proposed algorithm meets lower calibration data requirements, providing a new solution for the design of resource-constrained brain-computer interfaces. 2025 Vol. 38 (3): 280-292 [Abstract] ( 34 ) [HTML 1KB] [ PDF 717KB] ( 30 )

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