Neurocomputing最新文献

{"title":"MV-guided deformable convolution network for compressed video action recognition with P-frames","authors":"Yuting Mou,&nbsp;Ke Xu,&nbsp;Xinghao Jiang,&nbsp;Tanfeng Sun","doi":"10.1016/j.neucom.2025.130770","DOIUrl":"10.1016/j.neucom.2025.130770","url":null,"abstract":"<div><div>Large-scale deep models have driven substantial progress in action recognition, but their heavy computation and the use of full-resolution RGB frames raise latency and privacy concerns. Compressed-domain methods reduce overhead by operating on codec outputs (I-frames, P-frames) but still rely on privacy-sensitive I-frames and incur nontrivial decoding costs. To overcome these limitations, we propose a novel P-frame only framework that (1) employs deformable convolutions to exploit the spatial sparsity of residual maps in P-frames and (2) introduces a Motion Vector–guided Deformable Convolution Network (MV-DCN) that uses motion vectors to predict adaptive sampling offsets. To transfer semantic knowledge from RGB features without decoding I-frames, we further design a Motion-Appearance Mutual Learning (MA-ML) scheme for cross domain distillation. Extensive experiments demonstrate that our model achieves competitive accuracy and speed compared to raw domain and traditional compressed domain approaches, while effectively preserving privacy by utilizing only P-frames.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130770"},"PeriodicalIF":5.5,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144510718","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":"I2CKD : Intra- and inter-class knowledge distillation for semantic segmentation","authors":"Ayoub Karine ,&nbsp;Thibault Napoléon ,&nbsp;Maher Jridi","doi":"10.1016/j.neucom.2025.130791","DOIUrl":"10.1016/j.neucom.2025.130791","url":null,"abstract":"<div><div>This paper proposes a new knowledge distillation method tailored for image semantic segmentation, termed Intra- and Inter-Class Knowledge Distillation (I2CKD). The key novelty lies in its dual focus on transferring both intra-class and inter-class knowledge between intermediate layers of the teacher (cumbersome model) and student (compact model). For knowledge extraction, we exploit class prototypes derived from feature maps. To facilitate knowledge transfer, we employ a triplet loss in order to minimize intra-class variances and maximize inter-class variances between teacher and student prototypes. Consequently, I2CKD enables the student to better mimic the feature representation of the teacher for each class, thereby enhancing the segmentation performance of the compact network. Extensive experiments on four segmentation datasets, i.e., Cityscapes, Pascal VOC, CamVid and ADE20K, using various teacher–student network pairs demonstrate the effectiveness of the proposed method.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130791"},"PeriodicalIF":5.5,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144513826","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":"Multi-view Unsupervised Feature Selection via Global and Local Kernelized Graph Learning","authors":"Min Xu ,&nbsp;Xijiong Xie ,&nbsp;Yuqi Li ,&nbsp;Guoqing Chao","doi":"10.1016/j.neucom.2025.130786","DOIUrl":"10.1016/j.neucom.2025.130786","url":null,"abstract":"<div><div>To cope with the impact of the nonlinear characteristics of real data on feature selection, we propose a new multi-view unsupervised feature selection method called Multi-view Unsupervised Feature Selection via Global and Local Kernelized Graph Learning (FSGLK). Our method leverages the self-representation property of samples to capture global structures through kernelized graph learning and utilizes the learned kernelized graph to construct high-order tensors for capturing high-order relationships between views. In addition, we employ a kernelized adaptive neighborhood strategy to enhance the model’s ability to capture the local structures of complex data. The constructed graph can more effectively capture both the local and global structures of multi-view data while eliminating redundant features in high-dimensional data. Symmetric non-negative matrix factorization is used to obtain low-dimensional representations, on which feature selection is performed. To flexibly control matrix row sparsity, the <span><math><msub><mrow><mi>ℓ</mi></mrow><mrow><mn>2</mn><mo>,</mo><mi>p</mi></mrow></msub></math></span>-norm is introduced. Experimental evaluations on multiple benchmark datasets show that the proposed FSGLK method significantly outperforms existing methods in terms of clustering accuracy, consistency and information sharing.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130786"},"PeriodicalIF":5.5,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144502547","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 progressive sampling method for object detection performance surface based on Gaussian process multi-kernel fusion","authors":"Pengcheng Wang,&nbsp;Huanyu Liu,&nbsp;Junbao Li","doi":"10.1016/j.neucom.2025.130775","DOIUrl":"10.1016/j.neucom.2025.130775","url":null,"abstract":"<div><div>With increased interest in robustness evaluation of deep learning models, performance assessments under single-dimensional perturbations have been extensively studied, resulting in the establishment of numerous benchmarks. However, the behavior of models under bi-dimensional perturbations remains underexplored. A key issue arises from the exponential growth in sampling requirements when modeling performance surfaces in two-dimensional perturbation spaces, resulting in significant computational overhead. To address this issue, we propose a progressive sampling method for object detection performance surfaces that uses multi-kernel Gaussian process fusion. Our method incorporates a genetic algorithm to optimize kernel composition, leveraging the superior surface fitting capabilities and uncertainty quantification of composite kernel Gaussian processes. A reinforcement learning strategy is used to generate an initial population with high diversity and broad coverage. In addition, a whale optimization algorithm is used to fine-tune the weights and parameters of individual kernels, thereby improving sampling efficiency. Experimental results show that the proposed method significantly improves the sampling efficiency of performance surfaces, effectively reducing the number of samples required. This provides a reliable and efficient solution for robustness evaluation of deep learning models under complex perturbation scenarios.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130775"},"PeriodicalIF":5.5,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144523757","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":"VOS: Towards thermal infrared image colorization via View Overlap Strategy","authors":"Weida Zhan ,&nbsp;Yilin Wang ,&nbsp;Yu Chen ,&nbsp;Hang Yang ,&nbsp;Guilong Zhao ,&nbsp;Yingying Wang ,&nbsp;Shujie Zhai ,&nbsp;Tianyun Luan ,&nbsp;Deng Han","doi":"10.1016/j.neucom.2025.130793","DOIUrl":"10.1016/j.neucom.2025.130793","url":null,"abstract":"<div><div>Currently, a significant challenge still exists in thermal infrared images colorization, as current methods struggle with translating texture naturally and achieving color accuracy. To overcome this challenge, we propose a View Overlap Strategy (VOS) for colorizing infrared images. The proposed VOS employs a dual-branch generator designed to translate different regions of the same object into colorization output, and it evaluates the generated overlapping regions through an Optimal Adversarial Strategy (OAS) to determine the best generator output results. To achieve whole-image colorization, a unique sliding mechanism is designed that gradually extends the colorized region over the entire infrared image, continually approximating the final colorization result during the dual-branch generator’s adversarial training. Extensive experiments on the FLIR dataset and KAIST dataset demonstrate that the proposed VOS can be applied within existing colorization adversarial networks, leading to superior performance metrics and visual quality. The color images generated through our proposed VOS present enhanced clarity and realism.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130793"},"PeriodicalIF":5.5,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144502549","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 novel lightweight deep attention network for automated nuclei segmentation in histopathology images","authors":"Rashika Bagri,&nbsp;Ankit Rajpal,&nbsp;Naveen Kumar","doi":"10.1016/j.neucom.2025.130797","DOIUrl":"10.1016/j.neucom.2025.130797","url":null,"abstract":"<div><div>Cell nuclei segmentation in histopathological images is essential for developing computer-aided diagnostic (CAD) systems for cancer diagnosis and prognosis. However, this task remains challenging due to poor staining, low contrast, irregular shapes, and overlapping nuclei in histology tissue images. Traditional encoder–decoder-based architectures often struggle to capture fine-grained spatial details while also being computationally expensive. To address these challenges, we propose a novel lightweight deep attention network comprising three residual blocks in the encoder, a bottleneck block, and three wavelet-driven attention blocks in the decoder. The residual blocks used in the encoder effectively extract high-level features, while the bottleneck block captures global multi-resolution features. A newly introduced wavelet-driven attention block in the decoder leverage high-frequency two-dimensional discrete wavelet transform coefficients, that captures the finer edge-level details that are often lost during the encoding process. Evaluated on two publicly available datasets, PanNuke and TNBC, the proposed architecture achieved five-fold cross-validation Jaccard Index scores of <span><math><mrow><mn>74</mn><mo>.</mo><mn>47</mn><mo>±</mo><mn>0</mn><mo>.</mo><mn>74</mn></mrow></math></span> and <span><math><mrow><mn>71</mn><mo>.</mo><mn>21</mn><mo>±</mo><mn>0</mn><mo>.</mo><mn>82</mn></mrow></math></span>, respectively, at a 95% confidence level. The proposed architecture has significantly fewer trainable parameters and a smaller model size than existing architectures without compromising its performance. To further validate its efficacy, the model was tested on the MonuSeg dataset as an independent cohort, achieving a Jaccard Index score of <span><math><mrow><mn>64</mn><mo>.</mo><mn>17</mn><mo>±</mo><mn>1</mn><mo>.</mo><mn>75</mn></mrow></math></span>. Wilcoxon signed-rank and Scott-Knott ESD tests confirmed that the proposed architecture is statistically superior to existing models. Finally, Grad-CAM heatmaps revealed its superior focus on nuclei regions compared to conventional designs.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130797"},"PeriodicalIF":5.5,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144513972","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":"SFAG-DeepLabv3+: An automatic segmentation approach for coronary angiography images","authors":"Yinsheng Chen ,&nbsp;Ying Zhang ,&nbsp;Miaomiao Jiang ,&nbsp;Jiahao Li ,&nbsp;Xu Han ,&nbsp;Kun Sun ,&nbsp;Fan Wang ,&nbsp;Jinwei Tian ,&nbsp;Bo Yu","doi":"10.1016/j.neucom.2025.130781","DOIUrl":"10.1016/j.neucom.2025.130781","url":null,"abstract":"<div><div>Automated segmentation of coronary angiography images is highly significant for computer-aided diagnosis of coronary heart disease. However, existing segmentation methods suffer from the problem of poor segmentation results caused by insufficient extraction and fusion of the features of the complex topological structure of blood vessels. In view of this, this paper proposes an automated segmentation method for coronary angiography images based on SFAG-DeepLabv3+. This method utilizes the Swin Transformer network to screen coronary angiography images and proposes a Filtering Smoothing Equalization (FSE) image enhancement method to improve the quality of angiography images. Furthermore, this paper proposes an improved automatic segmentation network for coronary arteries based on the DeepLabv3+. In the encoder section, an Adaptive hybrid Dilated convolution and double Pooling (ADP) module is proposed to enhance the ability to extract topological features of coronary blood vessels. Between the encoder and decoder, a Gaussian Context Spatial Fusion (GCSF) module is proposed to reduce information loss during the compression and decompression of information from the encoder to the decoder. In the decoder section, bicubic interpolation upsampling is employed to improve the continuity of the segmented blood vessel topology. To validate the effectiveness of the proposed method, experiments were conducted using both the ARCADE public dataset and a self-constructed CSH dataset. Experimental results demonstrate that the method proposed in this paper can perform effective feature extraction, fusion and correction on coronary angiography images, achieving average Dice coefficients of 0.9249 on the CSH dataset and 0.9156 on the ARCADE dataset.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"650 ","pages":"Article 130781"},"PeriodicalIF":5.5,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144549449","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":"UdpTrace: Utility-enhanced differential privacy scheme for trajectory data publishing","authors":"Wei Sun,&nbsp;Kui Zhao,&nbsp;Gang Liang,&nbsp;Zhiwei Liang,&nbsp;Lingla Jiang","doi":"10.1016/j.neucom.2025.130785","DOIUrl":"10.1016/j.neucom.2025.130785","url":null,"abstract":"<div><div>Differential privacy is a popular method for preserving privacy in trajectory data publishing, allowing the utilization of user data while protecting sensitive information. However, trajectory data publishing under differential privacy often suffers from low utility and limits its value for meaningful analysis. In this paper, we propose UdpTrace, a utility-enhanced differential privacy scheme for trajectory data publishing. First, we introduce a method for discretizing geographic regions based on trajectory density. This method employs a coarse first-layer grid to enhance privacy protection and a fine second-layer grid to capture detailed trajectory information, thereby maintaining privacy while improving the application value of the data. Second, we build a semantic trip transfer matrix by aggregating the location label transfer probabilities and the cell transfers under the corresponding labels to maintain the semantic connection between the start and end regions. Third, we develop a novel sampling method to generate realistic intermediate trajectory points leveraging local transition patterns captured in the second-layer grid. Experimental results indicate that the proposed scheme effectively preserves privacy while significantly improving the utility of trajectory data.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130785"},"PeriodicalIF":5.5,"publicationDate":"2025-06-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144502546","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":"Comparison of distance and reinforcement-learning rules in social-influence models","authors":"Norberto M. Grzywacz","doi":"10.1016/j.neucom.2025.130870","DOIUrl":"10.1016/j.neucom.2025.130870","url":null,"abstract":"<div><div>Values are essential for decision-making in people and machines. When a decision is to be made, relevant information is obtained and then, the course of action that maximizes expected value is selected. For people, values have socio-cultural components, with individuals learning from each other. This learning leads the socio-cultural organization of values that includes grouping and polarization. To model this organization, sociologists and social psychologists use Agent-Based Social-Influence models. In standard form, these models use distance-based rules, that is, the degree by which a person influences another is a function of the distance between their values. In this article, we also introduce social-influence rules based on reinforcement learning. This is the mechanism that the brain and artificial intelligence use to learn to optimize values. We report computer simulations of the dynamics of multi-agent, Social-influence models, using either distance or reinforcement-learning rules. Our results show that both types of rules account for grouping and polarization. However, reinforcement-learning rules lead to a better accounting of the number of groups, their polarization, and the degree of individuality, that is, agents not belonging to clusters. Our simulations also reveal factors that influence these results. These factors include contrarians, the number of interacting agents, and the dimensional richness of the value space. Finally, our results with reinforcement-learning rules show complex dynamics of values, including group instability, tipping points, and phase transitions in spontaneous group formation. We discuss the consequences of these results for artificial-intelligence systems learning from each other through social-influence models.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130870"},"PeriodicalIF":5.5,"publicationDate":"2025-06-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144518746","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":"Unsupervised contrastive learning using out-of-distribution data for long-tailed dataset","authors":"Cuong Manh Hoang ,&nbsp;Yeejin Lee ,&nbsp;Byeongkeun Kang","doi":"10.1016/j.neucom.2025.130779","DOIUrl":"10.1016/j.neucom.2025.130779","url":null,"abstract":"<div><div>This work addresses the task of self-supervised learning (SSL) on a long-tailed dataset that aims to learn balanced and well-separated representations for downstream tasks such as image classification. This task is crucial because the real world contains numerous object categories, and their distributions are inherently imbalanced. Towards robust SSL on a class-imbalanced dataset, we investigate leveraging a network trained using unlabeled out-of-distribution (OOD) data that are prevalently available online. We first train a network using both in-domain (ID) and sampled OOD data by back-propagating the proposed pseudo semantic discrimination loss alongside a domain discrimination loss. The OOD data sampling and loss functions are designed to learn a balanced and well-separated embedding space. Subsequently, we further optimize the network on ID data by unsupervised contrastive learning while using the previously trained network as a guiding network. The guiding network is utilized to select positive/negative samples and to control the strengths of attractive/repulsive forces in contrastive learning. We also distil and transfer its embedding space to the training network to maintain balancedness and separability. Through experiments on four publicly available long-tailed datasets, we demonstrate that the proposed method outperforms previous state-of-the-art methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130779"},"PeriodicalIF":5.5,"publicationDate":"2025-06-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144513825","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}