Neurocomputing最新文献_第8页

DABLN：An intelligent classification network based on breadth-based learning for noisy redundant signals DABLN：一种基于宽度学习的噪声冗余信号智能分类网络

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-24 DOI: 10.1016/j.neucom.2025.131366

Haoran Liu , Haiyang Pan , Jinde Zheng , Rui Yuan , Jinyu Tong

{"title":"DABLN：An intelligent classification network based on breadth-based learning for noisy redundant signals","authors":"Haoran Liu , Haiyang Pan , Jinde Zheng , Rui Yuan , Jinyu Tong","doi":"10.1016/j.neucom.2025.131366","DOIUrl":"10.1016/j.neucom.2025.131366","url":null,"abstract":"<div><div>In the field of artificial intelligence recognition and classification, deep-based learning and breadth-based learning have attracted extensive attention and research as two major methods. However, deep-based learning requires a large amount of data for training, and the training process is relatively slow. In contrast, breadth-based learning is fast to train and works well with small samples of data. Unfortunately, breadth-based learning may be more susceptible to noise and redundant features. Therefore, a new Distributed Adversarial Broad Learning Network (DABLN) is proposed. In DABLN, the risk of overfitting in the random mapped features is minimized through introducing distributed constraint terms in the optimization problem, resulting in a stable model that eliminates redundant information. Furthermore, a pair of non-parallel adversarial hyperplanes is obtained in the random feature space, which can capture different aspects and variations in the data, thereby reducing the sensitivity of the model to noise and other redundant information. The different experimental results show that compared to the breadth-based methods, DABLN method has significant advantages in precision, Recall, F<sub>1</sub>-score, Kappa and Accuracy.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131366"},"PeriodicalIF":6.5,"publicationDate":"2025-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144896564","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}

引用次数: 0

Real-time human pose estimation and tracking on monocular videos: A systematic literature review 单目视频的实时人体姿态估计和跟踪：系统的文献综述

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-23 DOI: 10.1016/j.neucom.2025.131309

Yingying Chen , Zhenan Feng , Daniel Paes , Daniel Nilsson , Ruggiero Lovreglio

{"title":"Real-time human pose estimation and tracking on monocular videos: A systematic literature review","authors":"Yingying Chen , Zhenan Feng , Daniel Paes , Daniel Nilsson , Ruggiero Lovreglio","doi":"10.1016/j.neucom.2025.131309","DOIUrl":"10.1016/j.neucom.2025.131309","url":null,"abstract":"<div><div>Real-time human pose estimation and tracking on monocular videos is a fundamental task in computer vision with a wide range of applications. Recently, benefiting from deep learning-based methods, it has received impressive progress in performance. Although some works have reviewed and summarised the advancements in this field, few have specifically focused on real-time performance and monocular video-based solutions. The goal of this review is to bridge this gap by providing a comprehensive understanding of real-time monocular video-based human pose estimation and tracking, encompassing both 2D and 3D domains, as well as single-person and multi-person scenarios. To achieve this objective, this paper systematically reviews 68 papers published between 2014 and 2024 to answer six research questions. This review brings new insights into computational efficiency measures and hardware configurations of existing methods. Additionally, this review provides a deep discussion on trade-off strategies for accuracy and efficiency in real-time systems. Finally, this review highlights promising directions for future research and provides practical solutions for real-world applications.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"655 ","pages":"Article 131309"},"PeriodicalIF":6.5,"publicationDate":"2025-08-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144933400","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}

引用次数: 0

Age-at-death estimation from 3D bone models of the pubic symphysis by using a deep multi-view learning approach 利用深度多视图学习方法从耻骨联合三维骨模型估计死亡年龄

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-23 DOI: 10.1016/j.neucom.2025.131299

Alejandro Manzanares , Marilyn Bello , Fernando Navarro , Javier Irurita , Inmaculada Alemán , Pablo Mesejo , Óscar Cordón

{"title":"Age-at-death estimation from 3D bone models of the pubic symphysis by using a deep multi-view learning approach","authors":"Alejandro Manzanares , Marilyn Bello , Fernando Navarro , Javier Irurita , Inmaculada Alemán , Pablo Mesejo , Óscar Cordón","doi":"10.1016/j.neucom.2025.131299","DOIUrl":"10.1016/j.neucom.2025.131299","url":null,"abstract":"<div><div>Forensic anthropology is a sub-field of physical anthropology that analyzes human skeletal remains of medical and legal interest. Within it, estimating an individual’s biological profile plays a crucial role for forensic anthropologists and pathologists in identifying victims and documenting evidence. In particular, the estimation of the age-at-death of an individual is crucial to reduce the range of possible coincidences during the human identification process. Most existing approaches rely on the forensic expert’s subjective evaluation and do not consider relevant characteristics that are not detectable by traditional visual methods. This paper proposes a deep multi-view learning approach to automatically estimate an individual’s age-at-death from three 2D <em>panoramic views</em> generated from a 3D bone model of their pubic symphysis. These three images correspond to the three principal coordinate axes and consist of 3-channel images generated from the spatial distribution map, normals’ deviation map, and gradient normals’ deviation map projections. Our approach not only detects high-level characteristics of the pubic symphysis undetectable by the human eye, but also reduces the age estimation error to 6.74, becoming, as far as we know, the best result in the state-of-the-art. In addition, the proposed approach includes a post-hoc explanation stage that visualizes, through a 3D saliency map of the pubic symphysis, the most relevant pixel areas on which the proposed neural architecture relies to make a given decision. This provides forensic scientists with valuable information and confidence to support their decision-making process.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131299"},"PeriodicalIF":6.5,"publicationDate":"2025-08-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144911672","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}

引用次数: 0

PHLA: A pre-trained model based method incorporating HLA-peptide binding and immunogenicity PHLA：一种基于预训练模型的方法，结合hla肽结合和免疫原性

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-23 DOI: 10.1016/j.neucom.2025.131292

Jixiang Fan , Jianzhang Zhang , Ying Huang , Xiu-Xiu Zhan , Huajun Hu , Chuang Liu

{"title":"PHLA: A pre-trained model based method incorporating HLA-peptide binding and immunogenicity","authors":"Jixiang Fan , Jianzhang Zhang , Ying Huang , Xiu-Xiu Zhan , Huajun Hu , Chuang Liu","doi":"10.1016/j.neucom.2025.131292","DOIUrl":"10.1016/j.neucom.2025.131292","url":null,"abstract":"<div><div>The discovery of neoantigen peptides plays a vital role in tumor immunotherapy. Currently, the prediction of neoantigen peptides mainly considers the binding relationship between human leukocyte antigen (HLA) and peptides. However, by ignoring more comprehensive biological considerations, some results may not be sufficiently reliable in tumor immunotherapy. In this study, we applied natural language processing encoding technology and deep learning techniques for prediction, considering both the HLA-peptide binding probability (binding model) and the immunogenicity of the peptide-HLA complex (immune model). Experimental results show that the binding model of the peptide-HLA complex achieves better performance than some benchmarks on two external datasets. Meanwhile, the immunogenicity model of the peptide-HLA complex significantly improves the prediction performance. Finally, we perform data analysis on the prediction results and discover a close correlation with tumors, suggesting that the prediction results of the peptide-HLA complex may provide substantial assistance for tumor vaccine design.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131292"},"PeriodicalIF":6.5,"publicationDate":"2025-08-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908369","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}

引用次数: 0

SAGA-Feat: A semantic- and geometry-aware network for sparse local feature learning SAGA-Feat：一种用于稀疏局部特征学习的语义和几何感知网络

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-22 DOI: 10.1016/j.neucom.2025.131349

Yanhan Mo , Mengxiao Yin , Guiqing Li , Junjie Liao , Zhijie Liang

{"title":"SAGA-Feat: A semantic- and geometry-aware network for sparse local feature learning","authors":"Yanhan Mo , Mengxiao Yin , Guiqing Li , Junjie Liao , Zhijie Liang","doi":"10.1016/j.neucom.2025.131349","DOIUrl":"10.1016/j.neucom.2025.131349","url":null,"abstract":"<div><div>Sparse local feature learning plays a pivotal role in computer vision, underpinning applications such as simultaneous localization and mapping (SLAM), image matching, and 3D reconstruction. However, existing approaches often suffer from limited multi-scale contextual modeling, inadequate integration of spatial and semantic information, and significant detail loss during feature reconstruction. To address these challenges, we propose SAGA-Feat (Semantic- and Geometry-Aware Feature Extraction Network), a weakly supervised framework following a decoupled describe-then-detect paradigm. SAGA-Feat introduces a Semantic-and-Geometry-Aware Attention module (SAGA-Attn) to enhance directional sensitivity and local semantic aggregation, a Structure-Aware Fusion decoder (SAGA-Fuse) to enforce spatial–semantic consistency through dual-domain normalization, and a Semantic-Guided Adaptive Sampling module (SAGA-Sample) to refine feature reconstruction based on learned saliency cues. Extensive experiments across multiple benchmarks demonstrate that SAGA-Feat delivers state-of-the-art or competitive performance among learning-based methods, highlighting its overall robustness and effectiveness in complex visual scenarios.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"655 ","pages":"Article 131349"},"PeriodicalIF":6.5,"publicationDate":"2025-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144922337","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}

引用次数: 0

Efficient large margin adversarial training based on decision boundaries for adversarial robustness 基于对抗鲁棒性决策边界的高效大边际对抗训练

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-22 DOI: 10.1016/j.neucom.2025.131277

Mengjun Xu , Ziqiang Li , Lei Liu , Bin Li

{"title":"Efficient large margin adversarial training based on decision boundaries for adversarial robustness","authors":"Mengjun Xu , Ziqiang Li , Lei Liu , Bin Li","doi":"10.1016/j.neucom.2025.131277","DOIUrl":"10.1016/j.neucom.2025.131277","url":null,"abstract":"<div><div>Recent advances in adversarial training have been proposed to improve adversarial robustness by increasing feature margins, where robustness strongly correlates with improved input margins. However, we identify a fundamental limitation of these approaches in this work: increasing feature margins does not guarantee better input margins due to distance distortions in DNNs’ non-linear hierarchical structures, leading to limited robustness improvement. To address this challenge, we propose an <strong>E</strong>fficient <strong>L</strong>arge <strong>M</strong>argin <strong>A</strong>dversarial <strong>T</strong>raining method based on <strong>D</strong>ecision <strong>B</strong>oundaries (<strong>ELM-DB-AT</strong>). It reformulates input margin optimization as feature margin optimization by leveraging adversarial examples that minimally cross decision boundaries. Specifically, ELM-DB-AT first extracts input-margin-related features with pre-crafted misclassified examples that marginally cross the decision boundary. Then, it optimizes these features to achieve better adversarial robustness. This process establishes a crucial connection between feature margin optimization and input margin optimization. Furthermore, ELM-DB-AT regulates cross-entropy loss by combining three metrics (including intra-class feature compactness, inter-class feature separability, and feature margin), saving computational and time consumption. Extensive experiments demonstrate the effectiveness of our approach. For example, compared with the baseline methods using ResNet50 on CIFAR10, ELM-DB-AT improves robust accuracy and clean accuracy by up to 5.81 % and 5.18 %, respectively. Additional experiments on the ImageNette dataset further validate the superior capability of the proposed method.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131277"},"PeriodicalIF":6.5,"publicationDate":"2025-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908385","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}

引用次数: 0

A novel practically designated-time fuzzy optimal control via reinforcement learning for stochastic constrained nonlinear systems 基于强化学习的随机约束非线性系统实际指定时间模糊最优控制

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-22 DOI: 10.1016/j.neucom.2025.131315

Lifang Qiu , Junsheng Zhao , Zong-Yao Sun , Chaoxu Mu

{"title":"A novel practically designated-time fuzzy optimal control via reinforcement learning for stochastic constrained nonlinear systems","authors":"Lifang Qiu , Junsheng Zhao , Zong-Yao Sun , Chaoxu Mu","doi":"10.1016/j.neucom.2025.131315","DOIUrl":"10.1016/j.neucom.2025.131315","url":null,"abstract":"<div><div>This paper investigates the practically designated time fuzzy optimal control problem for stochastic constrained nonlinear systems with input delay, using reinforcement learning (RL) algorithms. For the uncertain stochastic nonlinear system, a new practically designated time control scheme is proposed and proved to be effective. Meanwhile, a fourth-order asymmetric time-varying barrier Lyapunov function (BLF) is leveraged to mitigate the severe uncertainty caused by the constraints, and Padé approximation and intermediate variables are deployed to counteract the impact of input delay. Additionally, the optimal RL-based instruction filtering control scheme is implemented by exploiting the designated time stability, which not only achieves a prescribed convergence time based on dynamic event triggering, but also ensures that the necessary state constraints are satisfied. The bounded command filtering design avoids the need to prove the continuity of the higher order derivatives of the virtual control signals, thus significantly reducing the complexity of the stability analysis associated with the designated-time convergence of the closed-loop signals. In the optimal control design, the controller put forward ensures that all signals within the closed-loop system remain bounded in probability and the tracking error meets the performance requirements. Finally, the simulation results of two examples verify the effectiveness and applicability.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131315"},"PeriodicalIF":6.5,"publicationDate":"2025-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908380","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}

引用次数: 0

Exponential synchronization of variable-order fractional spatio-temporal networks via impulsive control: A direct error technique 基于脉冲控制的变阶分数时空网络的指数同步：一种直接误差技术

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-22 DOI: 10.1016/j.neucom.2025.131272

Lingbi Tang , Cheng Hu , Juan Yu

引用次数: 0

RETO: Reinforcement learning enhanced terminology optimization for cyber threat intelligence summarization RETO：用于网络威胁情报总结的强化学习增强术语优化

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-22 DOI: 10.1016/j.neucom.2025.131339

Junmei Ding , Wei Peng , Yueming Lu , Zhiqiang Wang , Yan Feng

{"title":"RETO: Reinforcement learning enhanced terminology optimization for cyber threat intelligence summarization","authors":"Junmei Ding , Wei Peng , Yueming Lu , Zhiqiang Wang , Yan Feng","doi":"10.1016/j.neucom.2025.131339","DOIUrl":"10.1016/j.neucom.2025.131339","url":null,"abstract":"<div><div>Cyber threat intelligence summarization consolidates raw intelligence into concise summaries for decision-makers, playing a crucial role in rapid, accurate detection and response to cyber threats. However, existing research primarily focuses on improving the fluency, overlap, and accuracy of summaries, often neglecting the optimization of the correctness and coverage of specialized terminology. While recent advances in LLMs technology show promise, they frequently produce hallucinations, especially in specialized areas, undermining the reliability and accuracy of summaries. In this paper, a <strong>R</strong>einforcement Learning <strong>E</strong>nhanced <strong>T</strong>erminology <strong>O</strong>ptimization (<strong>RETO</strong>) approach is proposed for cyber threat intelligence summarization. <strong>RETO</strong>’s three modules include a two-stage terminology annotator, a threat intelligence-specific terminology generator, and a multi-reward-aware terminology optimizer. The first module provides high-quality annotated data, the second module trains an automatic terminology generator, and the final module optimizes the terminology generator, guiding LLMs to generate more accurate summaries within the reinforcement learning framework. Additionally, a multi-reward-aware function is designed to improve the terminology generation quality by maximizing a combination of ROUGE-L, Terminology Correctness, and Homologous Terminology Coverage. Experimental results on the CTISum dataset show that <strong>RETO</strong> achieves state-of-the-art performance in zero-shot summarization tasks, and the introduction of domain-specific terminology effectively reduces hallucinations generated by LLMs, leading to more accurate and reliable summaries. We also make our code publicly available at <span><span>https://github.com/JmeiDing/RETO</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131339"},"PeriodicalIF":6.5,"publicationDate":"2025-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908930","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}

引用次数: 0

DBAE: Decision boundary based adversarial example generation via multi-step optimization 基于多步优化的决策边界对抗示例生成

IF 6.5 2区计算机科学

Neurocomputing Pub Date : 2025-08-22 DOI: 10.1016/j.neucom.2025.131341

Yu Sun, Peiqi Sun, Yanhui Du

{"title":"DBAE: Decision boundary based adversarial example generation via multi-step optimization","authors":"Yu Sun, Peiqi Sun, Yanhui Du","doi":"10.1016/j.neucom.2025.131341","DOIUrl":"10.1016/j.neucom.2025.131341","url":null,"abstract":"<div><div>Adversarial examples are deceptive texts generated by introducing subtle perturbations into original samples, often imperceptible to humans. Existing adversarial example generation methods primarily focus on text modality, achieving high attack success rates but often at the cost of large perturbations, which compromise the quality of generated examples and may cause semantic drift. To address these limitations, we propose Decision Boundary-based Adversarial Example Generation (DBAE), a sentence-level method that leverages the decision boundary characteristics of the target model’s vector space. By optimizing the relative positioning of feature vectors to the decision boundary, DBAE generates high-quality adversarial examples that remain semantically close to the original text. Our approach begins with a Tree-based Autoencoder (Tree-AE), which generates initial adversarial example feature vectors by injecting perturbations into the root node of the text. Guided by these initial vectors, the Particle Swarm Optimization (PSO) algorithm searches for target feature vectors near the decision boundary while preserving semantic similarity. Finally, a Conditional LLM (c-LLM) maps the optimized feature vectors to high-quality adversarial examples. Experiments on multiple models and datasets demonstrate that DBAE not only maintains a high attack success rate but also improves perplexity and semantic similarity compared to traditional methods. Despite a slightly higher average generation time of 7.82 s on an RTX 4090, our decision-boundary method attains <span><math><mo>></mo><mn>85</mn></math></span> % semantic similarity and strong human-evaluation scores, confirming its effectiveness and practical value for adversarial text generation.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"654 ","pages":"Article 131341"},"PeriodicalIF":6.5,"publicationDate":"2025-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908933","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}

引用次数: 0