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....

2026 Vol.39 Issue.1, Published 2026-01-25 Papers and Reports    Researches and Applications        1 2026 Vol. 39 (1): 1-1 [Abstract] ( 24 ) [HTML 1KB] [ PDF 166KB] ( 23 ) Papers and Reports 2 Multi-window Multi-layer Perceptron Feature Pyramid Network for Unsupervised MRI Image Registration YU Han, SUN Zheng, ZHANG Shengnan, GAO Zhangshuo, DING Gang'ao Current unsupervised magnetic resonance imaging(MRI) registration methods are typically based on convolutional neural networks(CNNs) or Transformer architectures. However, significant limitations exist in both of them. CNNs have difficulty in modeling long-range dependencies because they are constrained by the local receptive fields. Transformers often struggle to achieve fine-grained registration at full image resolution due to the high computational complexity of the self-attention mechanism. To address these issues, a multi-window multi-layer perceptron feature pyramid network(PyraMLP-Net) is proposed. It is designed for efficient and accurate full-resolution brain MRI registration. First, a weight-sharing feature extraction module extracts multi-scale features from a pair of images through parallel dual-path convolutional encoding. Then, with the correlation-aware multi-window multi-layer perceptron being its core, a feature pyramid decoding module gradually fuses feature information of different scales through a bottom-up path to achieve coarse-to-fine optimization of the deformation field. Finally, a spatial transformation network module applies the deformation field as parameters to perform differentiable resampling on the image to be registered and generate the final registration result. Experiments on three public datasets demonstrate that PyraMLP-Net outperforms mainstream models in terms of registration accuracy, stability and efficiency. 2026 Vol. 39 (1): 2-30 [Abstract] ( 41 ) [HTML 1KB] [ PDF 8766KB] ( 53 ) 31 Dual-Branch Two-Stage Detection for Small Objects in UAV Images YANG Yi, ZHU Jiangrui, WANG Keping, ZHANG Gaopeng, QIAN Wei, WANG Tian The excessively small size of targets in images is a major challenge for drone-based object detection. Particularly when drones operate at high altitudes with low imaging resolution, features of small targets are prone to dissipation within the deep layers of deep neural networks. To address this issue, a method of dual-branch two-stage detection for small objects in unmanned aerial vehicle(DB-TS) is proposed. The parallel tasks consist of a small object detection task and a super-resolution reconstruction task. In the super-resolution reconstruction branch, a spatial prior module(SPM) and a window attention module(WAM) are constructed. The small object detection branch is built upon the Swin Transformer backbone. The spatial information from shallow features and the attention guidance from deep features are reconstructed via super-resolution methods of SPM and WAM, respectively. The two-stage detection framework consists of a training phase and an inference phase. During the training phase, the fine-grained detail extraction capability of the small object detection branch is strengthened by using the high-resolution features as ground truth in the super-resolution reconstruction branch. During the inference phase, inference speed is significantly improved and computational resource consumption is reduced by retaining only the small object detection branch.Experiments on VisDrone and JZ-UAV datasets demonstrate that the proposed method achieves higher recognition accuracy compared to baseline models and exhibits superior performance among compared state-of-the-art methods. 2026 Vol. 39 (1): 31-51 [Abstract] ( 28 ) [HTML 1KB] [ PDF 16590KB] ( 46 ) Researches and Applications 52 Global Multi-label Feature Selection Driven by Higher-Order Correlation and Dual Redundancy DENG Wen, SHE Yanhong, ZHENG Wenli, HE Xiaoli, QIAN Ting Multi-label feature selection is a critical preprocessing technique for handling high-dimensional multi-label data. However, existing approaches are often trapped in local optima due to greedy search strategies or unadequate measuring feature correlation and redundancy within sparse models. To address these issues, a global multi-label feature selection algorithm driven by higher-order correlation and dual redundancy(GHC-DR) is proposed. First, a fuzzy dependency measure based on multi-label k-nearest neighbors is introduced to accurately evaluate the higher-order correlations between features and the label system. Second, GHC-DR is designed to focus on the local geometric structure of features by constructing a feature graph to capture local similarities among features, and a dual redundancy evaluation mechanism fusing information theory with local structure is developed. Finally, higher-order correlation, dual redundancy and label correlations are integrated into a unified sparse learning objective function, and an efficient closed-form solution is derived. Experiments on 15 public multi-label benchmark datasets demonstrate the superior performance of GHC-DR across multiple evaluation metrics. 2026 Vol. 39 (1): 52-66 [Abstract] ( 26 ) [HTML 1KB] [ PDF 1266KB] ( 31 ) 67 High-Time-Step-Friendly Slice Parallel Spiking Neuron LAI Jianxiang, HUANG Fangwan, WU Yuezhong, YU Zhiyong Spiking neural networks are faced with challenges in time series prediction, including the difficulty of balancing computational efficiency and information processing capability in temporal mapping, insufficient scalability of parallel spiking neurons at high time steps, and a tendency to fall into local optimum. To address these issues, a high-time-step-friendly slice parallel spiking neuron(HSPSN) is proposed in this paper. First, a direct temporal mapping method is employed to achieve one-to-one matching of time steps. Then, a slice parallel mechanism is designed to integrate local and global temporal patterns at the neuronal level through the synergy of local and global slices. Finally, a constriction matrix random dropout strategy is adopted to effectively guide the neurons toward superior convergence. Experiments on seven real-world time series prediction datasets demonstrate that HSPSN significantly outperforms existing spiking neural networks in terms of prediction accuracy, energy efficiency and convergence stability and it can effectively capture complex spatiotemporal dependencies in multivariate time series and covariate-based time series. 2026 Vol. 39 (1): 67-82 [Abstract] ( 25 ) [HTML 1KB] [ PDF 979KB] ( 23 ) 83 Reversible Flow Network for Style Transfer Based on Frequency-Domain Enhanced Adaptive Channel Attention and Feature Pyramid Fusion GE Bin, SHAO Ziyi, ZHENG Junshuai, XIA Chenxing, GUAN Junming, XU Tao To address the issues of content distortion, artifact appearance, and insufficient utilization of frequency-domain characteristics in style transfer, a reversible flow network for style transfer based on frequency-domain enhanced adaptive channel attention and feature pyramid fusion(FECANet) is proposed. Based on the pre-trained VGG19 architecture, a reversible flow network is designed to reduce feature loss and ensure the integrity of content structure by leveraging its unbiased feature transfer mechanism. A frequency-domain enhanced adaptive channel attention module is developed to analyze the frequency-domain distribution of style images, and accurate correlations between content and style features are established to improve the stylization effect. Additionally, a feature pyramid fusion scheme is designed to align global style with local textures, enhancing the coordination of transfer results. Experiments on MS-COCO and WikiArt datasets show that FECANet effectively balances style transfer and content preservation, and it shows superior performance in content structure integrity, stylization effect and computational efficiency. 2026 Vol. 39 (1): 83-96 [Abstract] ( 37 ) [HTML 1KB] [ PDF 5237KB] ( 45 )

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