International Journal of Intelligent Systems最新文献

{"title":"A Machine Learning Framework for the Identification of Ion Channel Isoforms Via Stochastic Gating Signatures","authors":"Mustafa Demirtaş,&nbsp;Abdullah Aksoy,&nbsp;Enes Yiğit","doi":"10.1155/int/6767530","DOIUrl":"https://doi.org/10.1155/int/6767530","url":null,"abstract":"<p>Accurate and automated discrimination of ion channel isoforms from ionic current recordings is a critical challenge in biophysical analysis and drug discovery. In this study, a high-accuracy classification structure integrating time–frequency signal representations with machine learning techniques is proposed to address this problem. To effectively capture the nonstationary nature of ionic currents, short-time Fourier transform (STFT) is applied to the raw time-domain recordings of two representative sodium channel isoforms (Nav1.5 and Nav1.6). Through this transformation, 1D signals are converted into 2D time–frequency images. A dataset comprising 845 STFT-based images (420 for Nav1.5 and 425 for Nav1.6) is constructed for training and testing procedures. Various feature extraction strategies and machine learning classifier combinations are investigated. Experimental results demonstrate that local binary pattern (LBP) features provide superior discriminative capability for these representations. Specifically, an ensemble learning strategy combining bagged trees with LBP features achieved the highest and most consistent performance, yielding training, validation, and testing accuracy of 94.6%, 91.34%, and 94.49%, respectively. The results obtained reveal that the presented method provides a promising proof-of-concept framework for automated ion channel isoform classification, demonstrating the practical potential of STFT-based feature representations combined with ensemble learning under limited-data conditions.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-08-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/6767530","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148943461","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Multimodal Fusion and Supervised Contrastive Learning for Robust Document Image Classification","authors":"Anshul Kashyap,&nbsp;Meenakshi Mittal,&nbsp;Satwinder Singh","doi":"10.1155/int/4241437","DOIUrl":"https://doi.org/10.1155/int/4241437","url":null,"abstract":"<p>As various industries adapt to the digital transformation, their ability to effectively manage and evaluate vast amounts of digital documents has become essential to operational success. In this regard, automated and accurate classification of such documents plays a critical role. Traditionally, unimodal approaches to document image classification (DIC) show limited robustness when applied to complex and noisy images, constraining accuracy. In parallel, the growing demand for lightweight and resource-efficient models on edge devices and other resource-constrained environments has increased the need for computationally efficient solutions. This study addresses these limitations by proposing two distinct, independent approaches to DIC, designed for different organizational, industrial, and practical needs. The first approach is a multimodal late-fusion framework that combines visual and textual modalities of a document image to improve classification accuracy while reducing cross-modal error propagation. The second approach is an efficiency-oriented, lightweight framework, referred to as supervised contrastive semantic retrieval (SCSR). Instead of employing computationally heavy dense classification layers, this approach integrates supervised contrastive representation learning with semantic similarity retrieval to directly utilize the learned embedding space for final classification, providing a high-performance, resource-efficient alternative for DIC. Experimental results show that the multimodal late-fusion framework outperforms employed unimodal baselines and prior studies, achieving 93.06% and 92.00% accuracy on the RVL-CDIP and Tobacco-3482 datasets, respectively. Further, the alternative lightweight SCSR framework achieves competitive accuracies of 80.76% and 87.89% on the same benchmarks. Both frameworks were also evaluated on RealDoc-16, an independently collected primary dataset, where they achieved strong performance.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/4241437","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148849217","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"An Explainable Hybrid CRF–Weighted-Average Ensemble–LIME Framework for Aspect-Based Sentiment Analysis of Multilingual Hospital Reviews","authors":"Demmelash Abate,&nbsp;Nilay Mistry","doi":"10.1155/int/4928353","DOIUrl":"https://doi.org/10.1155/int/4928353","url":null,"abstract":"<p>Aspect-Based Sentiment Analysis (ABSA) has emerged as an important research area for extracting fine-grained insights from user-generated healthcare reviews. The existing methods, however, tend to have limited interpretability, poor aspect extraction performance and be ill-suited to multilingual healthcare data. This study proposes a framework for Explainable Aspect-Based Sentiment Analysis (EABSEA) to analyse the sentiment of hospital reviews using conditional random fields, Weighted Average Ensemble learning, and Local Interpretable Model-Agnostic Explanations (LIME). To test the framework, a large dataset of 105,637 Google reviews from 160 hospitals in India was used. The pipeline used for preprocessing included multilingual text normalisation, tokenisation, part-of-speech tagging, language detection and lemmatisation. The average F1-score of the best hyperparameter tuning using conditional random fields for aspect term extraction was 0.9583. Aspects were extracted and used to build aspect-aware feature representations for Naïve Bayes, K-Nearest Neighbours and Support Vector Machine classifiers. A weighted-average ensemble strategy has been used to fuse all the predictions. The experimental results show that the proposed WAE framework achieved an average F1-score of 0.9173 and an accuracy of 96.87% for aspect classification, which are better than the single machine learning and deep learning baselines, such as Logistic Regression, Random Forest, XGBoost, BERT-Tiny and LSTM. The proposed ensemble has an average F1-score of 0.8653 and 95.67% accuracy on sentiment classification. Additionally, LIME explanations helped improve model transparency by pinpointing the terms that influence the sentiment prediction across different aspects. The results demonstrate that the combination of conditional random fields aspect extraction, ensemble learning and XAI can be an effective and explainable approach for healthcare review analysis and service quality evaluation.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/4928353","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148784704","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"FRENet: Feature Relocation Network for Building Segmentation in Remote Sensing Images","authors":"Xuqing Zeng,&nbsp;Jieyuan Hu,&nbsp;Guanglian Li,&nbsp;Rongqing Wu,&nbsp;Haixing Song,&nbsp;Qiuyi Bai,&nbsp;Fuyun He","doi":"10.1155/int/5578308","DOIUrl":"https://doi.org/10.1155/int/5578308","url":null,"abstract":"<p>The challenge of locating target regions owing to the absence of spatial position information in image-level labels is resolved by using class activation maps (CAMs) for building localization. However, due to the intrinsic difference between classification and segmentation tasks, CAM is only able to focus on salient regions and often ignores subsignificant regions. This leads to incomplete internal regions and unclear boundaries of buildings extracted using weakly supervised models. Therefore, we propose an effective weakly supervised building extraction method, FRENet, to solve the above problems. So as to preserve building’s edge information, we employ dual-flow edge compensation module to extract various frequency components of the structure and add an auxiliary branch to provide additional guidance. Moreover, we design a feature sensing module to relocate the building area, which further guides the model to generate more complete CAM. To evaluate the effectiveness of FRENet, we conducted experiments on the Postdam and Vaihingen datasets, comparing it with seven mainstream weakly supervised models. FRENet outperforms AEFT by 2.93% and 4.47%, respectively, on the Postdam dataset, outperforming the competing approaches in terms of F1 score and mIoU. On the Vaihingen dataset, the F1 score and mIoU of FRENet reached 89.91% and 81.75%, respectively, which were 1.19% and 1.94% higher than the second-best ACGC. To further explore the source of this improvement, we visualized and analyzed frequency components of feature maps along with CAMs. The results provide insights into the feature learning advantages of FRENet and empirically validate the effectiveness of its design.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/5578308","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148784657","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"EDuHINet: An Educational Heterogeneous Information Network for Online Course Performance Prediction","authors":"Lanjun Luo,&nbsp;Peidong Long,&nbsp;Yuxuan Zhang,&nbsp;Ziqian Wan,&nbsp;Xueyan Wang","doi":"10.1155/int/6886082","DOIUrl":"https://doi.org/10.1155/int/6886082","url":null,"abstract":"<p>Accurate online course performance prediction is important for timely academic support, yet existing studies often rely on either static learner attributes or temporal behavioral data alone, without modeling their interaction explicitly. To address this limitation, this study proposes EDuHINet, an EDucational Heterogeneous Information Network that integrates cross-sectional static features with assessment-aligned temporal variables for multiclass online course performance prediction. Using 14,181 student records from OULAD, the model combines a temporal encoder, a static encoder, bilinear cross-modal interaction, and hierarchical fusion to capture both evolving learning behavior and stable learner context. The prediction target is defined as a four-category final course outcome classification, including Distinction, Pass, Fail, and Withdrawn. Experimental results show that EDuHINet achieves the strongest overall performance among the compared models, with advantages across overall accuracy, macro- and weighted-level F-score metrics, MCC, and Brier score, while maintaining competitive macro ROC AUC. Ablation experiments further indicate that joint modeling of the two modalities is more effective than using either modality alone, and early prediction results show that predictive information accumulates over the course trajectory. SHAP and ALE analyses show that late-stage assessment scores strongly drive model predictions. Meanwhile, static variables like precourse engagement and study load provide essential contextual support. These findings indicate that online course performance prediction benefits from an interpretable bimodal framework that captures both temporal progression and learner context.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-08-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/6886082","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148783549","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A Collective Attention Deep Reinforcement Learning Framework for Constrained Multidepot Vehicle Routing","authors":"Guangyu Zou,&nbsp;Xiaocao Hu,&nbsp;Levent Yilmaz","doi":"10.1155/int/1159050","DOIUrl":"https://doi.org/10.1155/int/1159050","url":null,"abstract":"<p>Delivery route planning can be formulated as a complex multidepot vehicle routing problem (VRP) with capacity constraints, pickup–delivery requirements, and time-window restrictions. Existing studies rarely address all these constraints simultaneously. This paper proposes a novel collective attention-based deep reinforcement learning framework to solve such constrained routing problems. The encoder adopts a node-pair representation scheme to capture the dependency between pickup and delivery locations. Furthermore, the integration of self-attention within a graph neural network (GNN) enhances the representation of node features. A unified decoder jointly models couriers and customer nodes, enabling the coordinated optimization of order assignment and route planning. Computational experiments on both synthetic unit-square instances and real-world road networks demonstrate that the proposed method outperforms classical heuristics and existing reinforcement learning approaches in terms of solution quality and computational efficiency.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/1159050","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148753734","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Benchmarking Commercial Speech Recognition and Multimodal Large Language Models on Dysarthric Speech: Severity-Stratified Baselines and Architecture-Specific Prompting Effects","authors":"Ali Alsayegh,&nbsp;Tariq Masood","doi":"10.1155/int/6065038","DOIUrl":"https://doi.org/10.1155/int/6065038","url":null,"abstract":"<p>Voice-based human-machine interaction has become a primary means of accessing intelligent systems, yet individuals with dysarthria are systematically excluded by persistent gaps in recognition accuracy. Whilst automatic speech recognition (ASR) achieves word error rates (WER) below 5% on typical speech, performance degrades sharply for dysarthric speakers; and although multimodal large language models (MLLMs) might compensate for acoustic degradation through contextual reasoning, their zero-shot behaviour on such speech remains uncharacterised. To address this, we evaluate eight commercial speech-to-text services on the TORGO dysarthric speech corpus, comprising four conventional ASR systems (AssemblyAI, Whisper large-v3, Deepgram Nova-3 and Nova-3 Medical) and four MLLM-based systems (GPT-4o, GPT-4o Mini, Gemini 2.5 Pro and Gemini 2.5 Flash), across lexical accuracy, semantic preservation and cost-latency trade-offs. Recognition degraded consistently with severity: mild dysarthria reached low single-digit WER, around 1% to 2% for the leading systems, approaching typical-speech benchmarks, whereas severe dysarthria exceeded 51% WER for every system, with the MLLMs offering no advantage over conventional ASR at this tier under default settings. A four-condition prompt ablation revealed strongly architecture-specific effects, whereby, for the OpenAI multimodal models, any verbatim-transcription prompt reduced severe-tier WER mainly by suppressing nontarget-language drift, lowering GPT-4o from 60.1% to 52.9% and GPT-4o Mini from 66.0% to roughly 55%, with the three prompt variants performing similarly; the Gemini models, which did not exhibit this failure mode to the same extent, showed no consistent benefit and sometimes degraded. The semantic metrics correlated strongly with WER and were largely redundant with it in aggregate but isolated a high-error tail in which communicative intent was partly preserved despite poor lexical accuracy. We offer these severity-stratified, per-speaker baselines as a reusable reference for evidence-based technology selection in assistive voice interface deployment, rather than as a claim of system superiority.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-08-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/6065038","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148752951","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Blockchain-Based AI Intelligent Systems in Healthcare: A Decentralized Federated Learning Framework With Zero Trust","authors":"Martin Wynn,&nbsp;Sepideh Mollajafari,&nbsp;Joe Cox,&nbsp;Mah Rukh,&nbsp;Md Hasibul Alam Ratul","doi":"10.1155/int/6908869","DOIUrl":"https://doi.org/10.1155/int/6908869","url":null,"abstract":"<p>Artificial intelligence (AI), blockchain (BC), and federated learning (FL) offer significant potential for strengthening data privacy, security, and decentralized decision-making in healthcare. However, healthcare AI systems remain vulnerable to adversarial threats, including data poisoning, Byzantine behavior, Sybil attacks, and unauthorized access. This article provides an integrative review of the pertinent literature from which a conceptual framework for exploring the relationship between BC, FL, and Zero Trust (ZT) is developed in the context of AI in healthcare. An experimental framework for a ZT approach based on BC and FL is then designed, tested, and evaluated. The framework combines strict identity verification, permissioned BC access control, immutable auditability, decentralized model training, and history-aware threat detection to enhance robustness while preserving data privacy. A series of simulations was conducted to evaluate the framework under poisoning, Byzantine, and Sybil attack scenarios, using <i>F</i>1-score as the primary model-performance metric and latency, gas cost, and throughput as BC-performance metrics. The results show that poisoning clients were consistently detected and excluded once adversarial behavior began, with only minor temporary performance degradation and subsequent recovery of global model performance. Under Byzantine attacks, the proposed history-aware trust filter improved the global <i>F</i>1-score across evaluated configurations, including an increase from 0.7762 under attack to 0.9117 with defense in one configuration. BC operations remained efficient, with latency generally between approximately 19 and 39 ms and throughput of 31.65 transactions per second, although upload operations incurred the highest gas cost. Sybil attack experiments further confirmed that cryptographic identity verification and permissioned access control prevented malicious identities from affecting the FL process, while attack cost increased linearly with no operational gain. The findings demonstrate that the proposed framework can effectively balance privacy preservation, adversarial robustness, auditability, and performance, making it a promising approach for secure healthcare systems.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/6908869","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148647858","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Dynamic Time-Varying W-Shaped and M-Shaped Transfer Function Based Feature Selection Method for Large-Scale Sparse Data","authors":"Zi-Rui Xu,&nbsp;Yu-Cai Wang,&nbsp;Wei-Zhong Sun,&nbsp;Jie-Sheng Wang,&nbsp;Shi Li,&nbsp;Si-Yu Jin,&nbsp;Yu-Wei Song,&nbsp;Yu-Liang Qi,&nbsp;Zhi-Guang Si","doi":"10.1155/int/2389650","DOIUrl":"https://doi.org/10.1155/int/2389650","url":null,"abstract":"<p>Large-scale sparse data are characterized by a massive feature dimensionality accompanied by significant sparsity in data elements. As a critical data preprocessing step, feature selection can effectively eliminate redundant and irrelevant information from the data, thereby substantially enhancing both the operational efficiency and predictive accuracy of models. To address the challenge of feature selection for large-scale sparse data, a binary black-winged kite optimization algorithm based on dynamic time-varying W-shaped and M-shaped transfer functions is proposed. The W-shaped and M-shaped transfer functions are designed, and a time-modulated factor is introduced to enable them to dynamically adjust along with the evolutionary process of the algorithm, thereby deriving the time-varying transfer functions. Furthermore, on this basis, the Euclidean Metric Factor (EMF) is constructed according to the population distribution state in the continuous space, and the Feature Coding Difference Factor (FCDF) is developed based on the binary feature differences of the population in the discrete space. By integrating EMF and FCDF into the time-varying transfer functions, the dynamic time-varying transfer functions are ultimately established. This integrated function can better adapt to the search requirements across all stages of the algorithm, ensuring that the population consistently maintains a reasonable search rhythm and direction. Multiple sets of simulation experiments were designed to verify the effectiveness of the proposed method in solving feature selection tasks. Firstly, DTD-W-BKA and DTD-M-BKA were determined as the optimal variants across 12 large-scale sparse datasets. Secondly, comparative experiments against other wrapper algorithms were conducted on real-world cancer gene expression datasets, demonstrating that DTD-W-BKA possesses a significant advantage. Additionally, the performance stability of the proposed method was evaluated under different classifiers, and a systematic ablation study was utilized to validate the independent contributions and collaborative synergy of each respective module.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/2389650","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148647857","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"RNFD: An Adaptive RGB-NIR Heterogeneous Fusion System for Robust Fire Detection and Precise Localization","authors":"Jinyang Dai,&nbsp;Qiuyang Cao,&nbsp;Xiaoning Song,&nbsp;Qixing Zhang","doi":"10.1155/int/7387105","DOIUrl":"https://doi.org/10.1155/int/7387105","url":null,"abstract":"<p>Recent advances in image-based detection have significantly accelerated the development and deployment of intelligent firefighting drones and robotic systems. However, high false alarm rates and inaccurate localization continue to hinder emergency response effectiveness. This paper proposes a high-precision RGB-NIR flame detection and localization (RNFD) system that leverages the complementary characteristics of RGB and near-infrared (NIR) imagery to ensure robust perception in complex environments. Specifically, our contributions are threefold: First, a novel RGB-NIR flame dataset is constructed to address the lack of publicly available benchmarks for multimodal fire detection and to support future research. Second, we propose RNAFNet, a flame detection network incorporating a self-developed adaptive feature fusion (AFF) module to enable the effective integration of RGB and NIR features, thereby enhancing detection accuracy. Finally, we introduced a unified high-precision localization framework. This involves utilizing the GA-RNAFNet algorithm to optimize the intrinsic parameters and distortion coefficients of heterogeneous binocular cameras and pioneers the application of deep learning to RGB-NIR binocular depth estimation. By combining RNAFNet-OTSU dynamic segmentation with CREStereo stereo matching, we achieve precise flame positioning. Experimental results demonstrate that the RNFD system reduces the false alarm rate to just 1.3% and achieves a mAP of 96.3%. The localization error is reduced to 6.9%, enabling accurate flame positioning at distances of up to 60 m. These results highlight the strong potential of low-cost RGB-NIR sensor fusion in enhancing fire perception and provide valuable insights for the development of next-generation fire detection systems.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-07-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/7387105","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148616920","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}