Applied Soft Computing最新文献

{"title":"A structure-guided framework for multimodal optimization leveraging chaotic search and persistence-based clustering","authors":"Xiang Meng ,&nbsp;Yan Pei","doi":"10.1016/j.asoc.2026.114687","DOIUrl":"10.1016/j.asoc.2026.114687","url":null,"abstract":"<div><div>Multimodal optimization remains a fundamental challenge due to the difficulty of preserving population diversity while efficiently locating multiple optima. To address this challenge, we propose the structure-guided multimodal optimization (SRS-GMO) framework, which explicitly models the evolutionary process as a structured search over topological regions in the decision space, characterized by a novel structural relevance score (SRS). This score integrates persistence, spatial distinctiveness, and recent evolutionary activity to adaptively allocate computational resources to promising niches while maintaining overall structural diversity. SRS-GMO employs chaotic evolution based on logistic maps to enhance global exploration and facilitate escaping from local optima. We instantiate the framework into two variants: SRS-GMO-SO for single objective multimodal optimization problems and SRS-GMO-MO for multi-objective multimodal optimization problems, respectively. Extensive experiments on the CEC2013 and MMF benchmark suites demonstrate that SRS-GMO achieves superior performance in peak coverage, convergence stability, and structural diversity preservation. We also provide a per-module computational complexity analysis and discuss practical accelerations for scalability. These results validate the effectiveness of the proposed framework in addressing the challenges of both single- and multi-objective multimodal optimization.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114687"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191524","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"The attention mechanism for intelligent text generation of traditional cultural symbols in artistic design","authors":"Xi Xu ,&nbsp;Da Guo","doi":"10.1016/j.asoc.2026.114733","DOIUrl":"10.1016/j.asoc.2026.114733","url":null,"abstract":"<div><div>Amid the digital wave, how to leverage artificial intelligence technologies to accurately interpret and reconstruct traditional cultural symbols has become a key issue in the field of artistic design. Existing image caption generation methods can identify the visual features of objects. But they often fail to capture the inherent cultural spirit and symbolic meanings when dealing with cultural symbols with profound historical connotations and abstract metaphors. This results in generated texts that remain at the level of superficial visual descriptions, lacking cultural depth. In response to this semantic gap, this study proposes an intelligent text generation model integrated with a cross-modal attention mechanism. The model constructs a dual-stream encoding architecture based on the Visual Geometry Group 19 (VGG19) and Bidirectional Encoder Representations from Transformers (BERT). By introducing multi-head self-attention and cross-modal attention modules, it achieves dynamic alignment between visual regions and cultural semantics. Different from traditional methods, the model can actively retrieve key visual clues from images according to the generated textual context. It accurately describes the physical forms of symbols and effectively interprets the cultural implications behind them. Experimental results show that this model achieves scores of 0.42 and 1.31 on the Bilingual Evaluation Understudy-4 (BLEU-4) and Consensus-based Image Description Evaluation (CIDEr) metrics, respectively, significantly outperforming benchmark models. It provides a new paradigm with both accuracy and cultural expressiveness for the digital inheritance and innovative design of traditional cultural symbols.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114733"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191556","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Hierarchical attentive feature refinement network with cross-resolution detail preservation for medical image segmentation","authors":"Xin Shu ,&nbsp;Anqi Shi ,&nbsp;Xiaofang Guo ,&nbsp;Yan Fan ,&nbsp;Xin Zhang","doi":"10.1016/j.asoc.2026.114760","DOIUrl":"10.1016/j.asoc.2026.114760","url":null,"abstract":"<div><div>Deep learning technology, particularly hierarchical encoder-decoder architectures, is developing rapidly in medical image segmentation, which aids in precise identification, localization, and quantification of anatomical structures and diseased areas in clinical diagnosis. However, most of the existing models face two critical challenges: (1) hierarchical semantic gaps caused by inconsistent receptive adaptation across different network depths; (2) gradual loss of fine details in traditional multi-scale fusion schemes, which disproportionately affects small-scale pathological features. Therefore, this paper proposes a Hierarchical Attentive Feature Refinement Network with Cross-Resolution Detail Preservation (HAFRN-CDP) to tackle these problems. Firstly, a Dual-Branch Pooling Fusion Encoder (DPFE) structure is designed to capture features of different scales and types to effectively alleviate information loss in the encoder. Secondly, a Fine Detail Retention (FDR) module is proposed to emphasize the high-resolution features in the encoder head, perceiving features at different scales and semantic levels. Thirdly, a novel Convolution Block (CB), which is composed of a Multi-Scale Residual Aggregation (MSRA) module and a Bi-Branch Global Attention (BGA) module, is presented to reduce the interference of redundant information caused by multi-scale strategy. Finally, in the decoder part, this paper proposes a Cross-Level Feature Fusion (CLFF) module to recover the details and structure by fusing features at different stages and scales. Extensive comparative experiments are conducted on three biomedical image datasets, and the results demonstrate that our HAFRN-CDP exhibits competitive performance compared to the state-of-the-art, with mDice of 92.87 % on JSUAH-Cerebellum and 91.77 % on 2018DSB, while also maintaining high mPrecision of 85.42 % and mRecall of 95.77 % on BUSI (malignant), confirming both its accuracy and robustness across diverse biomedical segmentation tasks.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114760"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191558","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Node importance identification method for urban rail transit networks based on improved mutual information considering inter-station cross-dependencies","authors":"Yanhui Yin ,&nbsp;Guanghong Hu ,&nbsp;Zhuo Huang ,&nbsp;Jihui Chen ,&nbsp;Wencheng Huang","doi":"10.1016/j.asoc.2026.114755","DOIUrl":"10.1016/j.asoc.2026.114755","url":null,"abstract":"<div><div>Identifying and protecting the key nodes of urban rail transit networks (URTN) is crucial for the normal operation of rail transit. However, most of the previous studies mainly relied on a single topological structure or passenger flow to identify important nodes, ignoring the inter-station cross-dependencies among the stations. In this paper, an Improved Mutual Information (IMI) is introduced to identify the node importance of URTN considering inter-station cross-dependencies. IMI emphasizes that the influence of a node will spread to the adjacent nodes within the effective range, and the importance of a node is the result of the combined influence of the adjacent nodes within that range. The comparative numerical researches have been conducted, the effectiveness and reliability of the IMI algorithm are validated from three aspects: network connectivity, node importance evaluation, and statistical analysis. Furthermore, under random and intentional attacks, we utilize the relative size of the maximal connected subgraph (RSMCS) and the relative network efficiency (RNE) before and after nodes are attacked to study the invulnerability of undirected unweighted networks of URTN. Additionally, we utilize the weighted RSMCS and the weighted RNE to study the invulnerability of bi-directional weighted URTN networks. Finally, A case study is performed with the Chengdu URTN serving as the context, the results show that, the IMI is suitable for networks with strong inter-station cross-dependencies (especially URTN), and can identify nodes with low degree but crucial due to their high inter-station cross-dependencies. It has higher discrimination ability for nodes ranked lower, and can capture stations with lower node degree but closely related to other important stations or important routes. Finally, suggestions for improving the performance and invulnerability of URTN are proposed from four aspects including expanding the capacity of the station, strengthening the connection between urban bus network and URTN, protecting important stations and routes with strong inter-station cross-dependencies through IMI algorithm, improving operation plans and alleviating peak passenger flow pressure.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114755"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191541","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Differential-game algorithm for resource optimization of wireless sensor network","authors":"Ning Yao ,&nbsp;Dong Liu ,&nbsp;Bai Chen ,&nbsp;Xiaochen Hao ,&nbsp;Rongrong Yin","doi":"10.1016/j.asoc.2026.114641","DOIUrl":"10.1016/j.asoc.2026.114641","url":null,"abstract":"<div><div>In response to the resource competition optimization problem in wireless sensor network, this paper first uses differential evolution algorithm to transform single-object multi-objective optimization into multi-object single-objective optimization, and combines game theory to design a new differential game algorithm. The algorithm conforms to the differential evolution framework on the outer layer and meets the requirements of game theory on the inner layer. And theoretically proved the low complexity and algorithm convergence. Then, the designed differential game algorithm was applied to the wireless sensor network to design the channel allocation algorithm. The convergence and effectiveness of the algorithm were verified through simulation. Finally, the differential game algorithm was discussed, and it was theoretically proved that the algorithm is also applicable to multi-objective optimization of multiple objects, providing a design idea for low complexity convergence multi-objective optimization algorithms.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114641"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039862","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"RADSCL: Representation augmentation integrated with dual supervised contrastive learning for low-resource text classification","authors":"Jun Zhang ,&nbsp;Ze Kuang ,&nbsp;Fanfan Shen ,&nbsp;Jianfeng He ,&nbsp;Rui Sun ,&nbsp;Yanxiang He","doi":"10.1016/j.asoc.2026.114646","DOIUrl":"10.1016/j.asoc.2026.114646","url":null,"abstract":"<div><div>Text classification in low-resource regime is a challenging task. Data augmentation techniques can significantly alleviate the issue of insufficient training samples in such environments by generating new samples. However, existing data augmentation methods have not yet effectively solved the problems of hard samples that are hard to classify and insufficient model generalization ability, which makes the performance of text classification in low-resource regime still has room for improvement. To this end, this paper proposes a method that fuses representational data augmentation and dual supervised contrast learning (RADSCL) in low-resource regime. Representational data augmentation first uses dynamic span-cutoff to remove context-independent words, which reduces the parameters needed for mixup and lowers the computational cost. Then the meaningless property of PAD word embeddings is utilized to perform weighted mixing with the cut-off text to reduce the importance of certain words in the text that affect classification. Through the representation augmentation method proposed in this paper, high-quality hard positive samples can be generated to optimize the decision boundary of the model. On this basis, in view of the current contrastive learning has achieved significant results in enhancing text representation, this paper constructs a dual supervised contrastive learning framework. The framework not only uses supervised contrastive learning to focus on the contrastive relationship between positive and negative samples, but also fully learns the potential semantic relationship between different categories of samples under the role of soft labels. Label distribution contrastive learning further utilizes the distribution information of soft labels based on supervised contrastive learning to impose effective constraints on the text representation, which effectively improves the performance of the text classification model. Multiple sets of experimental results show that the performance of the RADSCL model outperforms any model that incorporates other data augmentation and supervised contrastive learning models by a considerable margin on three benchmark datasets. In particular, the average accuracy of RADSCL is improved by 2.76 %, 2.66 %, 0.91 %, and 2.75 %, respectively, relative to the fusion models such as EDA+CE+SCL, Mixup+CE+SCL, ADV+CE+SCL, and AWD+CE+SCL.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114646"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039870","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Relation-aware heterogeneous graph network multi-modal predictive modeling of stock movements","authors":"Abdullah Ali Salamai","doi":"10.1016/j.asoc.2026.114703","DOIUrl":"10.1016/j.asoc.2026.114703","url":null,"abstract":"<div><div>The predictive modeling of movements in stock price information has been common but a challenging task for achieving sustainable management of stock marketplaces. This can be attributed to the importance of this information in evading market risks and improving financing decisions. This task becomes more challenging because of the availability of different modalities of stock information (i.e., stock prices, tweets, events, charts, etc.) and the presence of lead-lag relationships. In this regard, graph neural networks (GNNs) have recently achieved great improvements in providing effective modeling and analysis of stock relationships. However, multi-modal stock data evolve with dynamism in the form of heterogeneous graph topologies that limit the representation power of the existing GNNs. To this end, this study presents a novel graph intelligence framework called relation-aware heterogeneous graph network (RHGN) for efficient prediction of stock movements from multi-modal information by learning different forms of relational knowledge from heterogeneous graphs constructed for every trading day. In particular, the relational graph convolutions are presented to extract distinctive node-wise relational knowledge from the subgraph of each type of relationship. Then, the multi-relation knowledge passing module is designed to empower the network to model the dynamicity of relational knowledge throughout various types of relationships. Simultaneously, the edge relations are encoded into trainable temporal representations, from which edgewise relational knowledge is learned by capturing the contextual interactions among stock entities. In RHGN, an adversarial graph learning mechanism is introduced to adaptively augment and perturb the node representations based on the gradient information to make the model robust against small oscillations in input data. Experimentations on four public multi-modal stock movement datasets have validated the efficiency of the RHGN over the state-of-the-art.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114703"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080298","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Optimization method of process parameters based on the constraints of surface quality and geometric feature of elements during copying turning","authors":"Zhiheng Chen ,&nbsp;Ning Li ,&nbsp;Shiying Tu ,&nbsp;Kaimao Zeng ,&nbsp;Juan Lu ,&nbsp;Dan Chen","doi":"10.1016/j.asoc.2026.114711","DOIUrl":"10.1016/j.asoc.2026.114711","url":null,"abstract":"<div><div>Machining quality and performance are key indicators of CNC machine tool. To enhance the machining of non-circular curves and rotating parts on mid- to low-end CNC machines, this study proposes an optimization method for contour turning parameters based on surface quality and geometric feature constraints. The method integrates trajectory and parameter optimization. First, the Double-Heads Snake (DHS) algorithm generates a smooth G-code tool path comprising straight lines, convex arcs, and concave arcs with varying curvatures. Multi-objective optimization models are then developed for these elements to enhance machining quality and performance stability. Using Back Propagation Neural Networks (BPNN), the models predict surface roughness (<em>Ra</em>) and cutting force (<em>F</em>) based on process parameters and curvature radius, analyzing their influence mechanisms. An improved African Vulture multi-objective optimization algorithm (MOAVOA_improve) is proposed and validated against other algorithms using CEC2009 benchmarks, demonstrating superior performance. Surface quality and geometric constraints guide tool path segmentation, merging adjacent path segments with minimal curvature impact on <em>Ra</em> into single regions. Optimized process parameters are then determined for each region based on curvature and multi-objective solutions. The method is applied to non-circular rotating workpieces and compared with Mastercam and experience-based parameters. Results show significant improvements in machining quality, validating the method's effectiveness and practicality.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114711"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 multi-modal brain image fusion technique using nakagami imaging and intuitionistic fuzzy sets","authors":"K.G. Lavanya,&nbsp;P. Dhanalakshmi,&nbsp;M. Nandhini","doi":"10.1016/j.asoc.2026.114706","DOIUrl":"10.1016/j.asoc.2026.114706","url":null,"abstract":"<div><div>Medical image fusion is essential for consolidating images from diverse modalities into a single image, offering comprehensive information for diagnosis and analysis. With the rapid evolution of brain imaging technologies, researchers increasingly focus on refining fuzzy fusion techniques to utilize the full potential of these modalities. However, existing fuzzy fusion approaches often face three key challenges: difficulty in enhancing low-contrast images, uncertainty in selecting appropriate fuzzy membership functions, and redundancy in preserving important features. To address these shortcomings, a Nakagami distribution and intuitionistic fuzzy set-based multi-modal brain image fusion (NIFMBF) framework is proposed. This method integrates three core innovations: First, Nakagami imaging (NI) is employed to enhance low-contrast areas and reveal imperceptible lesions; second, a novel intuitionistic fuzzy generator (IFG) is designed to transform NI outputs into intuitionistic fuzzy images that effectively handle vagueness and enhance fine structural details; third, gray-level co-occurrence matrix (GLCM)-based contrast feature extraction guides the fusion process, enabling the retention of crucial information while reducing redundant data. Extensive experiments conducted on three benchmark datasets and compared against five state-of-the-art fusion techniques demonstrate the superior performance of the proposed NIFMBF method. Quantitatively, the method achieves an average improvement of 12–20% for every considered fusion metric. Additionally, the proposed framework reduces computational time by nearly 30% highlighting its efficiency. These findings, validated by both qualitative and quantitative evaluations, underscore the efficacy and potential of NIFMBF for medical image fusion.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114706"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080294","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Quantum-enhanced recurrent models for cognitive–motor assessment","authors":"Basma Jalloul,&nbsp;Bassem Bouaziz,&nbsp;Walid Mahdi","doi":"10.1016/j.asoc.2026.114696","DOIUrl":"10.1016/j.asoc.2026.114696","url":null,"abstract":"<div><div>Accurate assessment of cognitive and motor function is fundamental for the early detection of neurodegenerative and cerebrovascular conditions such as Mild Cognitive Impairment (MCI), stroke, and Parkinson’s disease. While clinical evaluations are often subjective, recent advances in wearable sensing and pose estimation have enabled objective gait analysis across these disorders. However, the inherent noise, artifacts, and inter-subject variability of real-world clinical data remain challenging for conventional deep learning models, which tend to overfit synthetic data and generalize poorly. In this work, we propose a hybrid soft-computing framework that integrates parameterized quantum circuits with recurrent neural networks (QE-RNNs) to enhance feature representation and robustness. Quantum-enhanced embeddings are executed on real quantum hardware and combined with classical temporal modeling to capture complex brain–motor dynamics. Experimental results across synthetic and clinical gait datasets show that, whereas classical RNNs lose significant accuracy in noisy environments, QE-RNNs maintain strong generalization, achieving up to 99.25% accuracy in stroke-related gait analysis and 81.43% in Parkinson’s motor-state classification. These findings highlight the potential of quantum-inspired soft computing for developing resilient, explainable, and clinically relevant tools for motor assessment.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114696"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080041","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}