Neurocomputing最新文献

{"title":"Spatiotemporal semantic structural representation learning for image sequence prediction","authors":"Yechao Xu ,&nbsp;Zhengxing Sun ,&nbsp;Qian Li ,&nbsp;Yunhan Sun ,&nbsp;Yi Wang","doi":"10.1016/j.neucom.2025.130159","DOIUrl":"10.1016/j.neucom.2025.130159","url":null,"abstract":"<div><div>Image sequence prediction is a fundamental task in computer vision that neural networks predict what happens in next frames given by a sequence of images. Despite remarkable progress in recent years, it is still challenging for predictive models to learn robust representations avoiding blurry objects, due to several aspects. (i) <em>Incomplete semantic structures.</em> Spatial structures including textures, positions and shapes of potential semantics are not comprehensively modeled, where existing methods directly learn from raw pixels or limited structural features within specific categories. (ii) <em>Absent structural correlation.</em> Correlations between images are commonly modeled by recurrent and convolutional architectures in this direction, while leaving correlations among explicit structures lack of explorations. (iii) <em>Non-selective structural fusion</em>. Existing fusion of structures in this task focuses on the concatenation of data or intermediate features, resulting features being equally treated and representative ones are ignored. In order to address above problems, we propose a spatiotemporal semantic structural representation learning pipeline in this study. (i) For comprehensively spatial structural modeling, this pipeline starts by SAM-based segmenting potential semantic objects, then providing spatial structures including textures, positions and shapes. (ii) To realize structural correlation modeling, temporal self-attention modules are exploited in this pipeline to extract compressed temporal features in corresponding structures. (iii) Finally for fusion of structural features, spatiotemporal semantic structural representations are accessed by integrating compressed features through cross-attention based fusion modules. Extensive experiments are conducted on both KITTI, KTH and Sthv2 datasets, showing superior performance of our model and especially better visual quality of semantic parts.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"637 ","pages":"Article 130159"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143739936","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"FCAnet: A novel feature fusion approach to EEG emotion recognition based on cross-attention networks","authors":"Mingjie Li,&nbsp;Heming Huang,&nbsp;Kedi Huang","doi":"10.1016/j.neucom.2025.130102","DOIUrl":"10.1016/j.neucom.2025.130102","url":null,"abstract":"<div><div>Recognition of human emotions from complex, diverse, and variable electroencephalogram (EEG) signals based on feature fusion strategies is gaining significant attention in affective computing. However, existing strategies employ a unified learning paradigm to integrate multiple features, and they lead to the neglect of critical local patterns and intricate relationships among features. Furthermore, most of models are generally designed to be complex and difficult to leverage and interpret in practical applications. To address these issues, FCAnet, a novel generic cross-attention-based feature fusion approach is prpopsed. FCAnet views the embedding space of multiple features as interactive patterns and utilizes cross-attention mechanisms to analyze both spatial correlations and discrepancy information simultaneously. Specifically, a dual-branch feature extraction module (DBFE) is first designed to effectively capture the 3D differential entropy (DE) and 3D power spectral density (PSD) feature maps of EEG. Secondly, a cross-attention feature fusion network (CAFFN) integrates a designed discrepancy information injection block (DIIB) with a common information injection block (CIIB) unit, facilitating significant interaction between different features. Finally, a time augmentation block (TAB) is employed to recover information lost from high-level representations, reusing discriminative temporal feature maps. Experimental results on the datasets DEAP, SEED, SEED-IV, and MPED demonstrate that the proposed FCANet outperforms state-of-the-art methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130102"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143785292","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Improving Chinese character representation with formation tree","authors":"Yang Hong ,&nbsp;Xiaojun Qiao ,&nbsp;Yinfei Li ,&nbsp;Rui Li ,&nbsp;Junsong Zhang","doi":"10.1016/j.neucom.2025.130098","DOIUrl":"10.1016/j.neucom.2025.130098","url":null,"abstract":"<div><div>Learning effective representations for Chinese characters presents unique challenges, primarily due to the vast number of characters and their continuous growth, necessitating models that can handle an expanding category space. Additionally, the inherent sparsity of character usage complicates the generalization of learned representations. Prior research has explored radical-based sequences to overcome these issues, achieving progress in recognizing unseen characters. However, these approaches fail to fully exploit the inherent tree structure of such sequences. To address these limitations and leverage established data properties, we propose Formation Tree-CLIP (FT-CLIP). FT-CLIP utilizes formation trees to represent characters and incorporates a dedicated tree encoder, significantly improving performance in both seen and unseen character recognition tasks. We further introduce masking for both character images and tree nodes, enabling efficient and effective training. This approach accelerates training significantly (by a factor of two or more) while enhancing accuracy. Extensive experiments show that processing characters through formation trees aligns better with their inherent properties than direct sequential methods, significantly enhancing the generality and usability of the representations.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130098"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143785317","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Contrastive learning-based fuzzy support vector machine","authors":"Yunlong Gao ,&nbsp;Junwen Jiang ,&nbsp;Jinyan Pan ,&nbsp;Bingjie Yuan ,&nbsp;Haifeng Zhang ,&nbsp;Qingyuan Zhu","doi":"10.1016/j.neucom.2025.130101","DOIUrl":"10.1016/j.neucom.2025.130101","url":null,"abstract":"<div><div>SVM faces significant challenges when dealing with noisy data, particularly in the context of complex data distributions and high levels of noise. Existing models, while offering some improvements in robustness, struggle with adapting to various noise types, handling extreme outliers, and overly relying on classification boundaries, which affects stability and classification accuracy. To address these limitations, this paper proposes a novel Fuzzy Support Vector Machine model based on contrastive learning, designed to enhance robustness to noise and improve generalization. The proposed model incorporates neighborhood structural information through contrastive learning, which refines the evaluation of slack variables and reduces the impact of noisy and outlier samples. Additionally, a redesigned fuzzy membership fusion mechanism is introduced, enabling more accurate handling of uncertain data. The experimental results on the benchmark dataset indicate that the average rankings of the accuracy and G-mean of CL-FSVM with linear and Gaussian kernels are below 2, which is superior to several excellent FSVM variants, demonstrating the effectiveness of this model.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"637 ","pages":"Article 130101"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143760267","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Optimal bipartite consensus for multi-agent systems using twin Q-learning deterministic policy gradient algorithm with adaptive learning rate","authors":"Lianghao Ji ,&nbsp;Jiali Song ,&nbsp;Cuijuan Zhang ,&nbsp;Shasha Yang ,&nbsp;Jun Li","doi":"10.1016/j.neucom.2025.130096","DOIUrl":"10.1016/j.neucom.2025.130096","url":null,"abstract":"<div><div>We investigate the optimal bipartite consensus control (OBCC) problem for multi-agent systems (MASs) over a signed network. Due to the improper cooperation-competition strength (CCS) among agents, the system may be unstable or even non-convergent. Recognizing the close relationship between CCS and the training of the critic network, we propose a twin Q-learning deterministic policy gradient algorithm with an adaptive learning rate (ALR-TQDPG). First, an adaptive learning rate formula is established based on the CCS and historical temporal difference (TD) error variations. The weights of two factors are dynamically adjusted using the weight equation as training progresses, then dynamically adjusting the update magnitude (i.e., learning rate) of critic network weights. Second, to solve the underestimation problem of Q-value, a twin Q-learning algorithm is adopted to improve system performance. The addition of experience replay and target network methods enhances algorithm stability. Lyapunov stability theory and functional analysis are utilized to ensure the ALR-TQDPG algorithm’s convergence. Finally, numerical simulations confirm that the suggested approach is effective.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130096"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143777431","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A stochastic gradient tracking algorithm with adaptive momentum for distributed optimization","authors":"Yantao Li ,&nbsp;Hanqing Hu ,&nbsp;Keke Zhang ,&nbsp;Qingguo Lü ,&nbsp;Shaojiang Deng ,&nbsp;Huaqing Li","doi":"10.1016/j.neucom.2025.130095","DOIUrl":"10.1016/j.neucom.2025.130095","url":null,"abstract":"<div><div>In this paper, we study distributed optimization problems where each node owns a local convex cost function calculated as the average of multiple constituent functions, and multiple nodes collaborate to minimize the finite sum of these local functions. Reviewing existing work, distributed optimization methods with adaptive momentum that consider reducing computation costs have not yet been explored. To this aim, we propose a gradient tracking stochastic distributed optimization algorithm with adaptive momentum, called GTSADAM. GTSADAM combines the distributed adaptive momentum method for faster convergence with the variance reduction mechanism to reduce computation costs. We provide a convergence analysis indicating that, under certain step size conditions, GTSADAM achieves linear convergence in the mean to the exact optimal solution when each constituent function is strongly convex and smooth. Moreover, GTSADAM maintains the acceleration efficiency of adaptive momentum while minimizing computation costs, which is confirmed by numerical simulations, and its performance is better than that of existing methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"637 ","pages":"Article 130095"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143739381","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Mask-Q attention network for flare removal","authors":"Zihao Li ,&nbsp;Junming Feng ,&nbsp;Siyao Hao ,&nbsp;Yuze Wang ,&nbsp;Weibang Bai","doi":"10.1016/j.neucom.2025.130100","DOIUrl":"10.1016/j.neucom.2025.130100","url":null,"abstract":"<div><div>Lens flare is a common optical phenomenon that is typically undesirable because it significantly degrades image quality, thus affecting certain visual tasks. The existing main methods using CNN-based models and transformer-based models aimed at removing lens flare, however, perform poorly in removing large-scale flares because they lack the inductive bias for spatial equivariance and the ability to capture both global and local features simultaneously. This deficiency prevents networks from quickly pinpointing the locations of flares and from fully learning the contextual content necessary for repairing contaminated areas, and they also struggle with restoring uncontaminated regions. In this paper, we introduce the Mask-Q Attention Network, a multi-scale framework designed to address the flare removal problem. Our approach focuses on extracting both global and local features, leveraging ResBlock for local feature extraction and Mask-Q Attention for capturing global features. In the Mask-Q Attention, we enhance the localization capability of the attention mechanism by integrating the flare mask with the query vector (Q) in the self-attention process. The flare mask is obtained by binarizing the initial image. This integration with the flare mask effectively resolves the issue of the lack of spatial equivariance in Transformer blocks, providing the network with prior knowledge of the flare locations. Extensive experiments demonstrate MQANet’s superiority, achieving minimum gains of 0.5% PSNR and 0.6% SSIM on synthetic datasets, alongside a 0.1% SSIM improvement on real-world data. It also performs well in terms of the LPIPS metric.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"637 ","pages":"Article 130100"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143760260","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"GDHS: An efficient hybrid sampling method for multi-class imbalanced data classification","authors":"Yuanting Yan ,&nbsp;Yan Lv ,&nbsp;Shuangyue Han ,&nbsp;Chengjin Yu ,&nbsp;Peng Zhou","doi":"10.1016/j.neucom.2025.130088","DOIUrl":"10.1016/j.neucom.2025.130088","url":null,"abstract":"<div><div>Multi-class imbalanced data has received increasing attention from the imbalanced learning community. Numerous methods have emerged in recent years, each of them exhibits superiority in certain scenarios. However, these methods usually confronted with the challenge of the complicated correlations among multiple classes, and the intractable data overlapping between classes poses additional difficulties for modeling the interrelations between classes. To this end, this paper proposes an efficient hybrid sampling method called GDHS for multi-class imbalanced data classification. It considers both the learning difficulty and the generalization potential of the data for minority oversampling. Moreover, it also considers the overlapping problems between classes, introducing majority cleaning strategies to enhance the visibility of each class as well as learning performance. To be specific, it proposes three cleaning strategies for handling the class overlapping problem: (1) self-inner class overlapping information based majority cleaning, (2) global overlapping information based majority cleaning and (3) balanced majority cleaning. Numerical experiments over 20 data sets demonstrate the superiority of the proposed method in terms of mGM and MAUC compared with 12 state-of-the-art methods. The implement of the proposed GDHS in programming language Python is available at <span><span>https://github.com/ytyancp/GDHS</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"637 ","pages":"Article 130088"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143760261","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"SRAD: A spatially-aware reconstruction network with anomaly suppression for multi-class anomaly detection","authors":"Shuyun Li ,&nbsp;Zhi Li ,&nbsp;Rongxiang Wang ,&nbsp;Weidong Wang ,&nbsp;Long Zheng ,&nbsp;Yu Lu","doi":"10.1016/j.neucom.2025.130085","DOIUrl":"10.1016/j.neucom.2025.130085","url":null,"abstract":"<div><div>Reconstruction-based methods have achieved remarkable outcomes in unsupervised image anomaly detection by training separate models for different categories. However, when it comes to a practical unified model, these methods often face the “identical shortcut” problem, where both normal and abnormal samples can be recovered well, leading to failure in anomaly detection. To address this problem, we propose a <strong>S</strong>patially-aware <strong>R</strong>econstruction network with anomaly suppression for multi-class <strong>A</strong>nomaly <strong>D</strong>etection (SRAD). Firstly, we propose a Spatially-aware Channel Convolutional (SCC) neural network, which replaces general convolution with channel convolution and incorporates a Spatial Information Fusion (SIF) block during encoding and decoding. The SIF block is proposed to allow the model to capture rich feature representations while avoiding overemphasizing details. Secondly, to prevent the model from learning identical reconstruction, we propose an Anomaly Suppression Feedback Learning (ASFL) strategy. The ASFL strategy encourages the model to accurately reconstruct normal samples while inhibiting the reconstruction of abnormal samples through a feedback mechanism. Experiments on MVTec AD, VisA and BTAD datasets demonstrate the clear superiority of SRAD compared to previous state-of-the-art unified models, e.g., achieving 98.2% I-AUROC and 97.6% P-AUROC on multi-class MVTec AD dataset. Furthermore, SRAD exhibits a high frame rate of 51 FPS on a 2080 Ti GPU, making it potential for practical applications.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"637 ","pages":"Article 130085"},"PeriodicalIF":5.5,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143767526","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Enhanced inter-camera person re-identification leveraging mixed-order relation-aware recurrent neural network","authors":"Vidhyalakshmi MK ,&nbsp;Bhuvanesh Unhelkar ,&nbsp;Pravin R. Kshirsagar ,&nbsp;R. Thiagarajan","doi":"10.1016/j.neucom.2025.130123","DOIUrl":"10.1016/j.neucom.2025.130123","url":null,"abstract":"<div><div>The person Re-Identification (Re-ID) requires a significant quantity of the costly label information, whereas unsupervised ones are still unable to provide satisfactory identification performance. These results in the poor scalability due to the requirement of the laborious data collection and annotation process in real-world Re-id applications. Unsupervised Re-ID techniques not require identity label data, but have significantly worse and inadequate model performance. In this paper, Enhanced Inter-Camera Person Re-identification leveraging Mixed-Order Relation-Aware Recurrent Neural Network (EICPR-MORRNN-TTAO) is proposed. The input images are collected from Market-1501, MSMT17, and Duke MTMC-reID datasets. Afterward, the input image is supplied to pre-processing. In preprocessing, Unsharp Structure Guided Filtering (USGF) is employed to enhance image quality. The pre-processed image is supplied to classification phase for Re-identifying the Inter-Camera Person as Same and Different utilizing Mixed-Order Relation-Aware Recurrent Neural Network (MORRNN). Generally, MORRNN does not adopt any optimization methods to determine the ideal parameters to assure accurate person Re-identification. Hence, Triangulation Topology Aggregation Optimizer (TTAO) is proposed to enhance the weight parameters of MORRNN. The EICPR-MORRNN-TTAO method is implemented in Python. The metrics, like Mean Average Precision (MAP), Cumulative Matching Characteristic (CMC), recall, Rank-1, Rank-10, Rank-20, Entropy, error rate, and Receiver Operating Characteristic (ROC) is considered. The EICPR-MORRNN-TTAO method attains 23.10 %, 27.54 % and 25.72 %, higher mAP, 21.48 %, 17.73 %, 25.32 % higher CMC and 20.98 %, 26.66 % and 16.32 % lower Error rate, are compared with existing techniques, like Intra-camera supervised Re-ID (ICSP-RI-PR), Offline-online associated camera-aware proxies for unsupervised Re-ID(OSC-UPRI-EIC), and Unsupervised Re-ID with stochastic training strategy (UPRI-EIC-STS) respectively.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130123"},"PeriodicalIF":5.5,"publicationDate":"2025-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143821342","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}