Neurocomputing最新文献

{"title":"SqSFill : Joint spatial and spectral learning for high-fidelity image inpainting","authors":"Zihao Zhang ,&nbsp;Feifan Cai ,&nbsp;Qin Zhou ,&nbsp;Youdong Ding","doi":"10.1016/j.neucom.2025.130414","DOIUrl":"10.1016/j.neucom.2025.130414","url":null,"abstract":"<div><div>Image inpainting has made significant progress due to recent advances in deep learning. However, most generative inpainting networks face challenges such as producing blurry results that lack high-frequency details or introducing inconsistent structures. To address these issues, we propose a novel transformer-based approach, SqSFill, which exploits rich information in both spatial and spectral domains. Specifically, SqSFill incorporates Rectified Frequency Feature Extractor (RecFFE) in the early layers of the network to capture fine-grained details by leveraging frequency information, guided by frequency loss. Moreover, we design a Scout Attention Block with linear complexity to replace vanilla self-attention, thereby effectively capturing long-range dependencies with lower computational cost. By integrating the RecFFE and Scout Attention Block, SqSFill is able to generate both coherent structures and sharp textures. Extensive experiments demonstrate the proposed SqSFill achieves superior results, outperforming previous state-of-the-art approaches with fewer parameters.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"645 ","pages":"Article 130414"},"PeriodicalIF":5.5,"publicationDate":"2025-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144099668","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":"Quality boost of tabular data synthesis using interpolative cumulative distribution function decoding and type-specific conditioner","authors":"Seungchan Roh,&nbsp;Seunghwan Song,&nbsp;Kwan-Yong Park,&nbsp;Byoung-mo Koo,&nbsp;Jun-Geol Baek","doi":"10.1016/j.neucom.2025.130484","DOIUrl":"10.1016/j.neucom.2025.130484","url":null,"abstract":"<div><div>Tabular data synthesis is an important research area in terms of privacy and data utilization. To enhance the utilization of tabular data, data synthesis techniques are extensively explored. The primary goal of tabular data synthesis is to generate high-quality data that preserve original insights while reducing the risk of data breaches. In this study, we propose a novel generative adversarial network (GAN) for quality boost of tabular data synthesis. The method of transforming continuous variables and correct conditioning for capturing dependencies between variables is considered a critical factor in determining data quality. Therefore, our proposed method uses interpolative cumulative distribution function (CDF) decoding for continuous columns and type-specific conditioner. Interpolative CDF decoding addresses a limitation of the inverse CDF method that restricts the diversity of synthetic data. In addition, the type-specific conditioner conditions the interdependencies between columns by integrating both discrete and continuous conditions. The introduction of conditional dependencies enables the generator to accurately capture complex dependencies between columns, thereby enhancing the fidelity of the synthetic data. The proposed framework, encompassing the interpolation in the decoding process and the generation method for conditions, serves to render synthetic data more realistic. A comprehensive evaluation on six datasets demonstrated that the proposed method is effective in terms of data quality, usability, and privacy level of the synthesized data. The source code is available at <span><span>https://github.com/rch1025/Tabular-GAN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"645 ","pages":"Article 130484"},"PeriodicalIF":5.5,"publicationDate":"2025-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144116660","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 contrastive learning-based heterogeneous dual-branch network for source camera identification","authors":"Zijuan Han ,&nbsp;Yang Yang ,&nbsp;Jinkai Zhang ,&nbsp;Yang Li ,&nbsp;Yunxia Liu ,&nbsp;Ngai-Fong Bonnie Law","doi":"10.1016/j.neucom.2025.130406","DOIUrl":"10.1016/j.neucom.2025.130406","url":null,"abstract":"<div><div>Source camera identification has been a significant focus in image forensics over the past decades. However, as camera model and instance related forensic features are weak compared to image contents, identification performance is far from satisfactory for practical applications. This paper introduces a novel contrastive learning strategy, aimed at enhancing the learning of camera fingerprints by leveraging the similarity between the two branches in a heterogeneous dual-branch network. Initially, a heterogeneous dual-branch feature extraction module is designed, employing two distinct strategies: noise residual estimation and progressive direct estimation, to independently extract forensic information. Contrastive learning is then utilized to enhance shared forensic features related to camera models between the two branches while filtering out irrelevant content residuals. During training, in addition to supervised classification loss, both spatial and frequency losses are applied to ensure the features consistency between the two branches, thereby enhancing the similarity of the features learned by both branches in the spatial and frequency domains. Drawing inspiration from the peak correlation energy metric commonly used in traditional methods, a frequency domain correlation loss is proposed. Extensive experimental results on the Dresden and Vision datasets demonstrate that the proposed method outperforms state-of-the-art approaches. Furthermore, it shows improved robustness against common preprocessing attacks such as JPEG recompression and image resizing, making it more suitable for real-world applications.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"645 ","pages":"Article 130406"},"PeriodicalIF":5.5,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144107695","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":"Robust fMRI time-varying functional connectivity analysis using multivariate swarm decomposition","authors":"Charalampos Lamprou ,&nbsp;Georgios Apostolidis ,&nbsp;Aamna Alshehhi ,&nbsp;Leontios J. Hadjileontiadis ,&nbsp;Mohamed L. Seghier","doi":"10.1016/j.neucom.2025.130404","DOIUrl":"10.1016/j.neucom.2025.130404","url":null,"abstract":"<div><div>Time-varying functional connectivity (TVFC) measured with functional MRI (fMRI) captures dynamic changes in statistical dependencies among regional time series, which can be studied with instantaneous phase synchronization analyses. Phase extraction requires narrow-banded resting-state fMRI (rs-fMRI) data typically extracted with conventional band-pass filtering or advanced mode decomposition techniques. However, filtering methods often struggle to eliminate noise effectively, require prior knowledge of cutoff frequencies, and fail to account for non-stationarity in the data. Likewise, existing mode decomposition techniques strongly depend on input parameters and are less reliable for multivariate analyses. Here, we introduce multivariate swarm decomposition (MSwD), a bio-inspired signal decomposition technique that combines the iterative nature of empirical methods with a robust mathematical foundation. Using synthetic signals and real rs-fMRI data from the Human Connectome Project and the Autism Brain Imaging Data Exchange I, we showed that MSwD-based PS (MSwD-PS) outperforms four state-of-the-art decomposition techniques in several key areas: (1) being more robust to input parameters and better at detecting true synchronizations, showing a 3–65% lower normalized root mean square error in simulated data and being 15.8-73.1% less prone to identifying short biologically implausible transitions between brain states, and (2) showing a reduced likelihood of false positives, being less affected by spurious synchronizations. Likewise, MSwD-informed functional connectivity analysis improved subject fingerprinting and autism spectrum disorder classification using graph neural networks. Overall, MSwD-PS can reduce the risk of false positives in TVFC, which could be extremely useful for processing rs-fMRI data with unknown ground truth in diverse clinical populations.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"642 ","pages":"Article 130404"},"PeriodicalIF":5.5,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144069459","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":"Adaptation and learning of spatio-temporal thresholds in spiking neural networks","authors":"Jiahui Fu ,&nbsp;Shuiping Gou ,&nbsp;Peizhao Wang ,&nbsp;Licheng Jiao ,&nbsp;Zhang Guo ,&nbsp;Jisheng Li ,&nbsp;Rong Liu","doi":"10.1016/j.neucom.2025.130423","DOIUrl":"10.1016/j.neucom.2025.130423","url":null,"abstract":"<div><div>Spiking neural networks (SNNs) have attracted substantial attention in recent years due to their brain-inspired and event-driven characteristics. To mimic the behavior of biological neurons, the neuron in SNNs generates spikes to transmit information across the network once its membrane potential surpasses a certain firing threshold. Due to model complexity and computational challenges, the threshold is often set as a fixed value, which limits the rich dynamical features of neurons and is inconsistent with the dynamic nature of thresholds observed in biological systems. Additionally, treating the threshold as an optimized parameter presents challenges in achieving convergence and maintaining stability. Therefore, we introduce a spatio-temporal adjustment strategy for the firing threshold. We propose a Learnable Temporal Factor (LTF) to dynamically adapt the threshold over time and an Adaptive Learnable Spatial Factor (ALSF) to spatially extend the threshold. By coupling these factors with the neuronal dynamics, we achieve a stronger spike coding capacity by utilizing more information in the generation of spikes. Our experiments show that the proposed method yields remarkable performance on both static and neuromorphic datasets. Our code is available at <span><span>github.com/gzxdu/ST-Thresholds-SNN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"644 ","pages":"Article 130423"},"PeriodicalIF":5.5,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144090648","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":"CIFFormer: A Contextual Information Flow Guided Transformer for colorectal polyp segmentation","authors":"Cunlu Xu ,&nbsp;Long Lin ,&nbsp;Bin Wang ,&nbsp;Jun Liu","doi":"10.1016/j.neucom.2025.130413","DOIUrl":"10.1016/j.neucom.2025.130413","url":null,"abstract":"<div><div>Automatic segmentation of polyps in endoscopic images plays a critical role in the early diagnosis of colorectal cancer. In recent years, Visual Transformers, especially pyramid vision transformers, have achieved remarkable strides and become dominating methods in polyp segmentation. However, due to the high resemblance between polyps and normal tissues in terms of size, appearance, color, and other aspects, the pyramid vision transformer methods still face the challenges of the representation of fine-grained details and identifying highly disguised polyps that could be pivotal in precise segmentation of colorectal polyp. To address these challenges, we propose a novel Contextual Information Flow Guided Transformer (CIFFormer) for colorectal polyp segmentation to reconstruct the architecture of a pyramid vision transformer via a contextual information flow design. Our proposed method utilizes a pyramid-structured encoder to obtain multi-resolution feature maps. To effectively capture the target object’s features at various levels of detail, from coarse-grained global information to fine-grained local information, we design a Global-Local Feature Synergy Fusion module (GLFS). GLFS adopts a progressive fusion strategy, first fusing the features of adjacent hierarchy and then gradually fusing across the hierarchy. This allows the model to utilize the features of different semantic levels better and avoid the information loss caused by direct fusion. In addition, we also introduce a Multi-Density Global-Local Residual Module (MDGL). The multi-density residual units within MDGL improve feature propagation and information flow. By employing both local and global residual learning, the model gains a better ability to capture detailed information at both global and local scales. The experimental results demonstrate that our CIFFormer model surpasses 17 benchmark models and achieves state-of-the-art performance on five popular datasets. Furthermore, our model exhibits remarkable performance on two video datasets as well. The source code of this work is available at <span><span>https://github.com/lonlin404/CIFFormer</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"644 ","pages":"Article 130413"},"PeriodicalIF":5.5,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144099451","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":"Mapping-to-Parameter nonlinear functional regression with Iterative Local B-spline knot placement","authors":"Chengdong Shi,&nbsp;Xiao-Jun Zeng,&nbsp;Ching-Hsun Tseng,&nbsp;Wei Zhao","doi":"10.1016/j.neucom.2025.130403","DOIUrl":"10.1016/j.neucom.2025.130403","url":null,"abstract":"<div><div>Many real-world phenomena are inherently continuous, yet traditionally represented as finite-dimensional vectors or matrices of discrete data points. Functional data analysis offers a natural paradigm by modeling observations as continuous functions, preserving intrinsic continuity and structural dependencies, thereby better capturing real-world dynamics and their underlying truth. However, functional modeling within infinite-dimensional spaces presents significant challenges due to its infinite degrees of freedom and computational complexity. These difficulties have led most studies on functional regression to focus on linear models, with general nonlinear approaches remaining underdeveloped. This paper introduces the Mapping-to-Parameter model, a simple yet effective approach for nonlinear functional regression. The key idea is straightforward: transform nonlinear functional regression problems from infinite-dimensional function spaces to finite-dimensional parameter spaces, where standard machine learning techniques can be readily applied. This transformation is accomplished by uniformly approximating all input or output functions using a common set of B-spline basis functions of any chosen order and representing each function by its vector of basis coefficients. For optimal approximation, we develop a novel Iterative Local Placement algorithm that adaptively distributes knots according to localized function complexity while providing theoretical guarantees on approximation error bounds. The performance of the proposed knot placement algorithm is shown to be robust and efficient in both single-function approximation and multiple-function approximation contexts. Through several real-world applications, the effectiveness and superiority of the Mapping-to-Parameter model are demonstrated in handling both function-on-scalar regression and function-on-function regression problems, consistently outperforming state-of-the-art methods including statistical functional models, neural network models, and functional neural networks.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"644 ","pages":"Article 130403"},"PeriodicalIF":5.5,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144099363","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 review of Zeroing neural network: Theory, algorithm and application","authors":"Shaoting Cao ,&nbsp;Jie Jin ,&nbsp;Daobing Zhang ,&nbsp;Chaoyang Chen","doi":"10.1016/j.neucom.2025.130425","DOIUrl":"10.1016/j.neucom.2025.130425","url":null,"abstract":"<div><div>As a powerful method for solving complex computational equations, neural networks have attracted widespread attention due to their unique advantages. However, due to the existence of time-varying problems in practical applications, traditional Gradient neural network (GNN) models may not be able to satisfy the requirements for accurately solving such problems, leading to the emergence of a dynamic system solution method - Zeroing neural network (ZNN), a dynamic system solver designed specifically for solving various time-varying mathematical problems and real-time control applications. ZNN eliminates errors through the use of dynamic differential equations, fundamentally overcoming the limitations of GNN in effectively converging for time-varying problems. Additionally, considering different application demands in practical scenarios and interference from noise in realistic environments, various robust ZNN models with different convergence properties have emerged. This paper will summarize the development of ZNN models in recent years from theoretical foundation, algorithm improvement and practical application aspects, and finally prospect the future research directions of ZNN models to provide researchers with a systematic reference.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"643 ","pages":"Article 130425"},"PeriodicalIF":5.5,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144071102","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":"Modeling of distributed parameter systems based on independent partial derivative-physics-informed neural network","authors":"Lijie Wang ,&nbsp;Yangshu Lin ,&nbsp;Xinrong Yan ,&nbsp;Yuhao Shao ,&nbsp;Zuhua Xu ,&nbsp;Chao Yang ,&nbsp;Haidong Fan ,&nbsp;Yurong Xie ,&nbsp;Chenghang Zheng","doi":"10.1016/j.neucom.2025.130396","DOIUrl":"10.1016/j.neucom.2025.130396","url":null,"abstract":"<div><div>Physics-informed neural networks (PINNs) have attracted considerable interest due to their capacity to incorporate established physical principles, thus facilitating efficient training with a limited amount of observed data. This approach ensures that the results of supervised learning are in accordance with the governing physical laws of the system, thereby enhancing the interpretability of models. However, existing research predominantly employs a single network architecture to simultaneously learn the information of the partial differential equations (PDEs). Such architectures typically require large network sizes to learn the system characteristics, which leads to prolonged training times and inefficiencies. This paper introduces a novel modeling framework, termed the independent partial derivative-physics-informed neural network (IPD-PINN). In contrast to conventional methodologies, IPD-PINN employs a network comprising multiple subnetworks, with each subnetwork dedicated to the learning of individual partial derivatives in the PDEs. The modular network structure allows for flexible and scalable adjustments to the size of each subnetwork, accommodating modeling tasks of varying complexities. Consequently, the entire network is capable of reducing the required training time while maintaining precise modeling accuracy. The efficacy of the proposed IPD-PINN method is demonstrated through two simulation case studies, which highlight its potential to improve the efficiency and accuracy of IPD-PINN modeling.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"642 ","pages":"Article 130396"},"PeriodicalIF":5.5,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144069536","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 dual-branch deep interaction network for multi-channel speech enhancement","authors":"Xiaoyu Lian,&nbsp;Nan Xia,&nbsp;Gaole Dai,&nbsp;Hongqin Yang","doi":"10.1016/j.neucom.2025.130412","DOIUrl":"10.1016/j.neucom.2025.130412","url":null,"abstract":"<div><div>Multi-channel speech enhancement removes noise and reverberation interference from noisy speech signals captured by microphone arrays. In this paper, we propose a dual-branch deep interaction network (DBDINet) for multi-channel speech enhancement, which complements the important features of both time domain and time–frequency domain in the speech signal. We design a waveform and complex spectrum interaction module (WCIM) to interact deeply with the information of two domains and propose an efficient Conformer (eConformer) as a transition layer of the network to improve network efficiency. We conducted extensive experiments on the synthetic AISHELL-1 dataset and the CHiME-3 dataset. The experimental results show that the proposed method achieves competitive performance on several metrics while maintaining lower computational complexity with faster inference speed than existing advanced methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"643 ","pages":"Article 130412"},"PeriodicalIF":5.5,"publicationDate":"2025-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144071103","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}