Neurocomputing最新文献

{"title":"Dual-decoder conditional diffusion model based on spatial-domain difference compensation pre-fusion for infrared and visible image fusion","authors":"Zhenjia Zhang,&nbsp;Kai Huo,&nbsp;Jun Li,&nbsp;Shuaifeng Zhi","doi":"10.1016/j.neucom.2025.131342","DOIUrl":"10.1016/j.neucom.2025.131342","url":null,"abstract":"<div><div>Infrared and visible image fusion aims to preserve both intensity and texture details from the two modalities. While generative adversarial networks (GANs) have shown promise, they suffer from mode collapse and limited interpretability. To address these challenges, we propose a dual-decoder conditional diffusion model based on spatially compensated pre-fusion, termed SDC-DDF. This method tackles the lack of prior knowledge in fusion distribution and reduces the excessive inference steps typical of diffusion models. By leveraging spatial discrepancies between the two modalities, pre-fused images are generated as noisy \"fused ground truth\", serving as coarse prior in the forward diffusion process. Additionally the refinement decoder refines the pre-fused images with varying levels of noise. During the reverse denoising, coarse reconstruction decoder restores a coarse fused image from random noise in just five denoising steps, which is then refined by the refinement decoder to produce a high-quality final result. Experimental results on five publicly available datasets demonstrate that SDC-DDF outperforms eleven state-of-the-art methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131342"},"PeriodicalIF":6.5,"publicationDate":"2025-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144911801","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":"Fusing spatio-temporal information using supervised local low-rank correlation embedding for depression recognition","authors":"Lu Zhang ,&nbsp;Peng Xu ,&nbsp;Zhijun Yao ,&nbsp;Xinyan Zhang ,&nbsp;Juan Wang ,&nbsp;Bin Hu ,&nbsp;Gang Feng ,&nbsp;Hong Peng","doi":"10.1016/j.neucom.2025.131350","DOIUrl":"10.1016/j.neucom.2025.131350","url":null,"abstract":"<div><div>Electroencephalography (EEG) signals contain rich spatio-temporal information that reflects the brain’s dynamic activity, making it widely used in depression recognition. However, effectively integrating this information to capture discriminative and complementary features remains a key challenge. To address this issue, we propose a novel Discriminative Local Low-Rank Correlation Embedding (DLLCE) to fuse spatio-temporal information of EEG. DLLCE integrates shared low-rank representation, local invariance, discriminative constraints, and enhanced correlation analysis into a unified framework. Specifically, the shared low-rank representation is used to capture the common structural patterns, while the correlation analysis aims to reduce redundancy among feature sets. In addition, the Laplacian regularization is applied to the shared representation to preserve the local geometric structure of the original data. To further enhance discriminative capability, a discriminant graph embedding term is incorporated to exploit label information. Experimental results on EEG datasets demonstrate that DLLCE achieves superior performance compared to existing methods. This work provides new insights into EEG-based mental health assessment and holds promise for early depression diagnosis and clinical decision support.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"656 ","pages":"Article 131350"},"PeriodicalIF":6.5,"publicationDate":"2025-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145048457","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-source adversarial domain adaptation with modulated adaptive weights","authors":"Wenjun Liu,&nbsp;Na Chen,&nbsp;Jiangtao Peng,&nbsp;Shuoshuo Hui","doi":"10.1016/j.neucom.2025.131352","DOIUrl":"10.1016/j.neucom.2025.131352","url":null,"abstract":"<div><div>Domain drift, characterized by the distributional discrepancy between labeled source domain training data and unlabeled target domain testing data, poses a significant challenge for deep neural networks in cross-domain tasks. This challenge is exacerbated when handling data from multiple source domains. Multi-source unsupervised domain adaptation (MUDA) aims to address this issue by leveraging information from multiple sources to enhance model performance on the target domain. Despite notable advancements, existing MUDA methods still exhibit several limitations. Currently, feature discriminability receives an excessive penalty when using adversarial learning to learn transferable features with cross-domain knowledge in multi-source domain adaptation. To address this, we introduce batch spectral penalization (BSP), which enhances the relative strength of other feature vectors, thereby improving feature discriminability. Moreover, most of the existing methods assume that uniform contributions from all source domains, neglecting the inherent differences among them. Accordingly, we consider quantifying the discrepancies between each source domain and the target domain. Ultimately, we introduce multi-source adversarial domain adaptation with modulated adaptive weights (MADA-AW). This framework employs an adversarial learning strategy with an adaptive weight mechanism to dynamically assess and optimize the contribution of each source domain. Experimental results on multiple image classification benchmarks validate the effectiveness of MADA-AW.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"655 ","pages":"Article 131352"},"PeriodicalIF":6.5,"publicationDate":"2025-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145004478","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":"StyleMorpheus: Learning a StyleGAN-based 3D-aware morphable face model with a disentangled style space","authors":"Peizhi Yan ,&nbsp;Rabab K. Ward ,&nbsp;Dan Wang ,&nbsp;Qiang Tang ,&nbsp;Shan Du","doi":"10.1016/j.neucom.2025.131329","DOIUrl":"10.1016/j.neucom.2025.131329","url":null,"abstract":"<div><div>Recent advances in 3D-aware neural rendering have enabled photorealistic face image synthesis from arbitrary viewpoints. However, achieving disentangled control over facial attributes typically depends on large, curated datasets collected in controlled environments. To overcome this limitation, we introduce StyleMorpheus, a 3D-aware, StyleGAN-based morphable face model that can be trained entirely on in-the-wild face images. StyleMorpheus surpasses traditional 3D-aware morphable models in rendering quality, despite relying solely on unconstrained 2D training data. Unlike conventional StyleGAN-based methods, StyleMorpheus also provides disentangled control over facial identity, expression, and appearance, allowing each attribute to be adjusted independently without unintended changes to the others. StyleMorpheus employs an auto-encoder structure, where the encoder learns a representative, disentangled style code space, and the decoder enforces disentanglement by using shape- and appearance-related codes at different levels of the network. Furthermore, we fine-tune the decoder through StyleGAN-based generative adversarial learning to achieve photorealistic rendering quality. StyleMorpheus is computationally lightweight and achieves real-time rendering speeds, making it suitable for virtual reality applications. We further demonstrate the disentanglement capabilities of StyleMorpheus through face editing tasks such as style mixing, face morphing, and color editing. Project homepage: <span><span>https://peizhiyan.github.io/docs/morpheus</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131329"},"PeriodicalIF":6.5,"publicationDate":"2025-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144896538","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":"Text-graph alignment based on LLMs instruction learning and wavelet graph transformers for brain disease analysis","authors":"Wenfeng Tan ,&nbsp;Hong Yang ,&nbsp;Shanshan Ye ,&nbsp;Ji Zhang ,&nbsp;Peng Zhang ,&nbsp;Zhi Qiao","doi":"10.1016/j.neucom.2025.131355","DOIUrl":"10.1016/j.neucom.2025.131355","url":null,"abstract":"<div><div>Brain graphs from functional magnetic resonance imaging (fMRI) combined with phenotypic texts serve as fundamental data modalities for brain disease analysis. The lack of alignment between self-supervised signals in the graph modalities of brain graphs and the text modalities of phenotypic data in current multimodal learning methods has led to suboptimal diagnostic performance and limited interpretability of brain disease patterns. To address these limitations, we introduce the Brain Text–Graph Alignment Model (BrainTGA), which integrates brain disease diagnosis with pathological analysis. Our model creates a shared semantic space that aligns text and graph data by using Large Language Models (LLMs) instruction learning to enhance phenotypic texts and implementing a wavelet graph Transformer model to encode brain graphs. Our comprehensive evaluation of benchmark datasets with heterogeneity and site bias demonstrates that BrainTGA outperforms existing state-of-the-art methods in diagnostic accuracy. It also has advantages in terms of time and space complexity and provides superior interpretability of neural patterns consistent with the diagnoses of clinical experts. The open source model implementation of our BrainTGA is available at <span><span>https://github.com/Mmurphyy/BrainTGA</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131355"},"PeriodicalIF":6.5,"publicationDate":"2025-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908384","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":"F-ADMD: Fast anomaly detection via multi-scale guided denoising diffusion model","authors":"Congying An ,&nbsp;Jingjing Wu ,&nbsp;Huanlong Zhang","doi":"10.1016/j.neucom.2025.131305","DOIUrl":"10.1016/j.neucom.2025.131305","url":null,"abstract":"<div><div>Visual anomaly detection has become a significant solution in industrial production due to its remarkable effectiveness and efficiency. However, traditional generative model-based methods are constrained in overall performance due to limitations in reconstruction quality and speed. Recently, generative approaches based on diffusion models have brought new opportunities to visual tasks, owing to their high-quality and diverse generation capabilities. Therefore, this paper proposes a novel anomaly detection framework based on diffusion models, termed F-ADMD, which is primarily composed of two parts: image reconstruction and defect segmentation. In the image reconstruction portion, we introduce a multi-scale guided rapid reconstruction method that is not only significantly faster than traditional diffusion models but also effectively handles the reconstruction of various anomaly-type regions. In the defect segmentation section, we integrate a Transformer attention mechanism that can capture both local and global context information at different feature levels, aiding the model in focusing more effectively on key regions of the image and thus improving the accuracy of identifying the boundaries of anomalous areas and enhancing segmentation precision. F-ADMD achieves state-of-the-art performance in image-level detection and anomaly localization on the challenging and widely used MVTec dataset, demonstrating the framework's effectiveness and broad applicability. This new method not only enhances detection accuracy but also significantly boosts processing speed, providing a novel comprehensive solution for visual anomaly detection in industrial production.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131305"},"PeriodicalIF":6.5,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144894816","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":"SAGM-Net: Spatio-temporal action graph modeling network for weakly-supervised temporal action localization","authors":"Yutong Li ,&nbsp;Miao Ma ,&nbsp;Chao Yao ,&nbsp;Zhao Pei ,&nbsp;Jie Wu","doi":"10.1016/j.neucom.2025.131304","DOIUrl":"10.1016/j.neucom.2025.131304","url":null,"abstract":"<div><div>Weakly-supervised temporal action localization (WS-TAL) is a challenging task in computer vision that aims to identify and locate specific action instances in untrimmed videos with only video-level annotations. To achieve this task, some existing methods build upon multiple instance learning (MIL) frameworks or design feature enhancement modules based on pre-trained extracted features. However, due to the insufficient exploration of snippet-level contextual relationships, these methods still suffer from imprecise temporal localization. Furthermore, some approaches leverage the advantage of graph convolution networks in relational modeling to address these issues. Nonetheless, they overlook the differences in spatio-temporal relation graphs generated by different modalities, confusing some action snippets with ambiguous information. To this end, we propose a novel weakly-supervised temporal action localization method, namely spatio-temporal action graph modeling network (SAGM-Net). On the one hand, we design a semantic-to-structural graph modeling (SSGM) module to refine feature representation by leveraging temporal dependencies and semantic discriminability of different modalities. On the other hand, we introduce a separation-aware feature enhancement (SFE) module, which enhances the semantic representation capability of features and maximizes the separation between pseudo-action and pseudo-background features. Experiments conducted on THUMOS14 and ActivityNet v1.2 demonstrate that SAGM-Net outperforms most WS-TAL methods, achieving an average mAP of 45.9 % on THUMOS14 and 26.5 % on ActivityNet v1.2, with performance comparable to the AICL method (AAAI 2023) on THUMOS14 and the LPR method (TCSVT 2024) on ActivityNet v1.2.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131304"},"PeriodicalIF":6.5,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144894833","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":"SAFCN: Self-enhancing attention fusion contrastive network for multimodal sentiment analysis","authors":"Yijun Chen ,&nbsp;Lina Yang ,&nbsp;Thomas Wu ,&nbsp;Yifeng Tan ,&nbsp;Xichun Li ,&nbsp;Zhonghua Li ,&nbsp;Wanxing Zhu","doi":"10.1016/j.neucom.2025.131212","DOIUrl":"10.1016/j.neucom.2025.131212","url":null,"abstract":"<div><div>Multimodal sentiment analysis (MSA) aims to analyze users’ sentiment information through heterogeneous data modalities. While contemporary MSA methodologies predominantly employ transformer-based architectures and related techniques for sentiment prediction, they neglect a comprehensive exploration of both unimodal feature refinement and cross-modal representation synthesis. To address these limitations, we propose the Self-Enhancing Attention Fusion Contrastive Network (SAFCN), a novel framework integrating hierarchical attention mechanisms with contrastive learning paradigms to achieve enhanced multimodal fusion. Specifically, three self-augmented encoders independently encode visual, textual, and auditory features, thereby generating optimized unimodal representations. Subsequently, a Cross-Modal Reinforced Fusion (CRF) module facilitates inter-modal interactions to derive enriched cross-modal representations. These representations are fused via a variety of fusion methods, with multimodal fusion further assisted by contrastive learning. Experimental results on three public datasets demonstrate the effectiveness of the proposed method. On the CH-SIMS dataset, SAFCN outperforms current state-of-the-art methods, achieving improvements in Acc-2, F1, Acc-3, MAE, and Corr of <span><math><mn>0.74</mn><mspace></mspace><mi>%</mi></math></span>, <span><math><mn>0.31</mn><mspace></mspace><mi>%</mi></math></span>, <span><math><mn>0.7</mn><mspace></mspace><mi>%</mi></math></span>, <span><math><mn>0.013</mn></math></span>, and <span><math><mn>0.041</mn></math></span>, respectively. On CMU-MOSI and CMU-MOSEI, SAFCN delivers state-of-the-art comparable performance, with notable advancements: on CMU-MOSI, it surpasses previous methods by <span><math><mn>3.8</mn><mspace></mspace><mi>%</mi></math></span> in Acc-7, <span><math><mn>0.002</mn></math></span> in MAE, and 0.04 in Corr; on CMU-MOSEI, it achieves a <span><math><mn>0.09</mn><mspace></mspace><mi>%</mi></math></span> improvement in F1-score.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131212"},"PeriodicalIF":6.5,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908931","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":"Global attractivity criteria for a discrete-time Hopfield neural network model with unbounded delays via singular M—matrices","authors":"José J. Oliveira ,&nbsp;Ana Sofia Teixeira","doi":"10.1016/j.neucom.2025.131311","DOIUrl":"10.1016/j.neucom.2025.131311","url":null,"abstract":"<div><div>In this work, we establish two global attractivity criteria for a multidimensional discrete-time non-autonomous Hopfield neural network model with infinite delays and delays in the leakage terms. The first criterion, which applies when the activation functions are bounded, is based on <span><math><mi>M</mi></math></span>-matrices that are not necessarily invertible. The second criterion, relevant for unbounded activation functions, requires that a related singular <span><math><mi>M</mi></math></span>-matrix be irreducible. We contrast our findings with existing results in the literature and present numerical simulations to illustrate the novelty of the proposed criteria.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131311"},"PeriodicalIF":6.5,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144903310","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 measure of conditional dependence for multivariate compositional data","authors":"Keyao Wang ,&nbsp;Huiwen Wang ,&nbsp;Shan Lu ,&nbsp;Lihong Wang","doi":"10.1016/j.neucom.2025.131331","DOIUrl":"10.1016/j.neucom.2025.131331","url":null,"abstract":"<div><div>Recent advances in predictive modeling of multivariate compositional data, which encompass scenarios with multiple compositional predictors and either compositional or scalar responses, have highlighted critical needs for effective dependence measurement. Current methodologies face limitations in quantifying conditional dependencies between responses and individual predictors while controlling for confounding variables, particularly when addressing nonlinear associations. To bridge this gap, we present a novel Compositional Conditional Dependence Coefficient (CoCDC) featuring four key innovations. (1) We introduce a nearest neighbor analysis designed for compositional predictors, and establish a coefficient for measuring conditional dependence between scalar responses and multivariate compositional predictors. (2) Through optimal transport theory, we propose a multivariate rank for compositions with formal definitions and estimation procedures. (3) We further extend CoCDC to handle predictions involving both compositional responses and predictors through integration of the proposed multivariate rank for compositional responses and the nearest neighbor analysis for compositional predictors. (4) We propose a Feature Ordering by Compositional Conditional Dependence (FOCCD) algorithm that leverages the CoCDC to enable effective variable selection. Extensive simulations reveal that CoCDC achieves competitive performance in detecting both linear and nonlinear marginal associations compared to benchmarks, while establishing the first measurement for quantifying joint dependencies in compositional predictions with either compositional or scalar responses. The FOCCD algorithm demonstrates superior performance in complex prediction scenarios involving nonmonotonic marginal associations and joint dependencies, outperforming existing methods in variable selection accuracy. Empirical studies through real-world datasets further confirm the superiority of CoCDC, with FOCCD-selected predictors yielding parsimonious and interpretable models.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131331"},"PeriodicalIF":6.5,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908386","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}