Procedia Computer Science最新文献

{"title":"A Resource-Efficient Codebook-Driven Semantic Structuring Pipeline for Human-AI Dialogue in Ambient Intelligent Systems","authors":"Aisvarya Adeseye ,&nbsp;Jouni Isoaho ,&nbsp;Seppo Virtanen ,&nbsp;Tahir Mohammad","doi":"10.1016/j.procs.2026.04.070","DOIUrl":"10.1016/j.procs.2026.04.070","url":null,"abstract":"<div><div>Human–AI dialogue in ambient intelligent systems is increasingly relying on large language models (LLMs). When questions are generated dynamically to enable personalized and context-aware interactions, variations in phrasing and topical focus exist between conversations. Without structured organization, which is often extremely resource-intensive, conversational data remains fragmented and cannot be reliably used for systematic analysis or reporting. This study proposes a semantic structuring pipeline to map LLM-generated questions to shared codes, sub-themes, and themes using a predefined codebook. This multi-stage pipeline applies semantic screening, factor-based scoring, mathematical aggregation, and validation checks, supported by locally deployed LLMs and manual confirmation. The pipeline was evaluated on 6,030 question–response pairs collected from dynamic interviews across three research objectives. The framework achieved an overall mapping accuracy of 97% while reducing hallucinated semantic matches to 1.2% through layered validation. The results indicate that the framework effectively reduces hallucinated matches and improves mapping accuracy while remaining computationally efficient for private local deployment.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"280 ","pages":"Pages 552-559"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141062","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Low-Latency Visual Product Search in E-Commerce Using CNN Models","authors":"Esther M. Matendo ,&nbsp;Selain K. Kasereka ,&nbsp;Alidor M. Mbayandjambe ,&nbsp;Rostin M. Mabela ,&nbsp;Alain M. Kuyunsa ,&nbsp;Jean Didier M. Batubenga ,&nbsp;Tasho Tashev ,&nbsp;Kyandoghere Kyamakya","doi":"10.1016/j.procs.2026.04.087","DOIUrl":"10.1016/j.procs.2026.04.087","url":null,"abstract":"<div><div>Visual product search has gained attention as an alternative to text-based queries, yet most existing approaches emphasize accuracy or scalability while giving limited consideration to end-to-end latency in deployment settings. This work presents a deployment-oriented evaluation of a two-phase visual product search pipeline that combines MobileNetV3-Large for feature extraction with FAISS HNSW for efficient similarity indexing. The offline phase embeds product images into 512-dimensional vectors, while the online phase performs low-latency query processing. Real-time constraints are achieved under GPU inference, whereas CPU-only deployment represents a conservative baseline that exceeds the target latency. Experiments conducted on 44,000 fashion products spanning seven categories demonstrate strong performance. The model achieves 99.77% overall Classification accuracy, with 99.86% for Apparel (14,923 samples), 99.88% for Footwear (6,417 samples), and 99.61% for Accessories (7,882 samples). Under class-imbalanced conditions, evaluation using weighted metrics yields a weighted F1-score of 0.9977, reflecting consistent performance across dominant categories. For visual retrieval, FAISS HNSW delivers sub-millisecond search latency (0.29 ms) with high similarity scores. The end-to-end pipeline processes queries in 71.58 ms on CPU hardware, dominated by CNN inference and preprocessing, while the similarity search contributes negligible overhead. Although the CPU-based implementation exceeds the 30 ms real-time target, clear optimization pathways exist through GPU acceleration and INT8 quantization. Overall, this study provides practical insights into deploying low-latency visual search systems for e-commerce applications, balancing accuracy, scalability, and computational efficiency.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"280 ","pages":"Pages 687-694"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141066","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Embarak’s 5-Pillars for Adaptive Learning: An AI-Driven Framework for Early Detection and Personalized Intervention for At-Risk Students","authors":"Ossama Embarak ,&nbsp;Adel Zairi","doi":"10.1016/j.procs.2026.04.088","DOIUrl":"10.1016/j.procs.2026.04.088","url":null,"abstract":"<div><div>Early identification of at-risk students and timely personalized interventions are critical for improving academic outcomes and student retention. Traditional predictive approaches often rely primarily on academic performance indicators, overlooking behavioral, contextual, and engagement-related factors, which limits their effectiveness. This paper proposes Embarak’s 5-Pillars Adaptive Learning (E5P-AL) framework, an AI-driven and holistic approach that integrates predictive modeling, generative AI, and human-guided oversight to deliver personalized interventions across five key dimensions: academic performance, attendance and participation, engagement and learning behavior, contextual and socioeconomic factors, and continuous monitoring and feedback. Using a multi-modal dataset comprising 1,000 student profiles, E5P-AL demonstrates superior predictive performance compared to baseline models, achieving an accuracy of 0.94, an F1-score of 0.92, and an AUC of 0.97. Ablation experiments confirm that the joint contribution of all five pillars—particularly engagement and contextual dimensions—is essential for optimal prediction accuracy. Furthermore, the framework’s continuous performance tracking and adaptive intervention mechanisms led to a 35% reduction in predicted risk levels and a 28% increase in student engagement over a four-week period. Overall, E5P-AL provides a scalable, modular, and ethically grounded solution for adaptive learning across diverse educational contexts.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"280 ","pages":"Pages 695-704"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141067","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Real-Time Route Recommendation Framework for Congested Urban Networks","authors":"Gamil Ahmed ,&nbsp;Tarek Sheltami ,&nbsp;Ashraf Mahmoud ,&nbsp;Mustafa Ghaleb","doi":"10.1016/j.procs.2026.04.092","DOIUrl":"10.1016/j.procs.2026.04.092","url":null,"abstract":"<div><div>Efficient route guidance in congested cities requires optimizing travel time, because signal timing, localized bottlenecks, and time-varying congestion can make longer routes faster than the shortest path. This paper presents a real-time route recommendation framework for Thessaloniki that integrates road network topology with spatiotemporal traffic speeds to support both shortest-distance and fastest-time routing. The framework preprocesses urban road data to construct distance- and time-weighted graphs, enabling dynamic route selection under peak and off-peak conditions. Using real-world mobility and traffic datasets, the proposed approach consistently recommends congestion-aware routes that can be physically longer yet significantly faster during heavy traffic. Experimental results across multiple scenarios demonstrate travel-time reductions of up to 88% during peak periods, while maintaining route optimality with respect to the selected objective. A comparative evaluation of A* and Dijkstra’s algorithm shows that A* achieves lower execution times, making it more suitable for real-time navigation in dense urban networks.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"280 ","pages":"Pages 729-736"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141160","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Intelligent Classification Model of Terrain-Level Real-Scene 3D Point Clouds Based on Deep Learning","authors":"Jie He ,&nbsp;Liang Liao ,&nbsp;Yu Lei ,&nbsp;Yongzhi Lin ,&nbsp;Yan Gao","doi":"10.1016/j.procs.2026.04.193","DOIUrl":"10.1016/j.procs.2026.04.193","url":null,"abstract":"<div><div>Rapid change identification of terrain-level real-scene 3D point cloud data is difficult, and the lack of high-precision classification methods leads to low DEM (Digital Elevation Model) update efficiency, necessitating an intelligent point cloud classification model based on deep learning. This paper firstly identifies potential terrain change areas by combining voxel grid downsampling with change detection, providing candidate points for deep learning classification; then, an improved Point Transformer network is constructed, integrating self-attention mechanism and multi-scale features to achieve high-precision intelligent classification of complex terrain point clouds; finally, semantic consistency optimization and noise removal are performed on the classification results, and the DEM is updated based on the classification results, realizing automated high-precision digital elevation model reconstruction after terrain changes. Experiments show that the proposed model achieves classification accuracy of 0.90 for buildings and 0.91 for water bodies, with an overall accuracy exceeding 0.9, thus solving the problem of high-precision point cloud classification. After DEM update, most errors are concentrated within ±0.3m, with an overall average error of 0.099m, significantly improving the accuracy and efficiency of DEM updates.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"281 ","pages":"Pages 65-73"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141172","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Big Data-Driven Development of Precise Exercise Energy Consumption Calculation and Personalized Health Management Algorithms","authors":"Xiaopeng Lin,&nbsp;Li Li","doi":"10.1016/j.procs.2026.04.195","DOIUrl":"10.1016/j.procs.2026.04.195","url":null,"abstract":"<div><div>As the volume of basketball training and game data grows rapidly, the current approaches to estimating the energy consumption of sports and various health-related issues include low precision of the round level energy consumption, lack of individual differences, and the inability to respond to the real-time risk of training in case of danger. To overcome this, this paper presents a big data-based algorithm that determines the energy consumption of basketball-related sports and offers individual health control. This algorithm uses a combination of wearable sensors, multi-view court video paths and physiological to create a three-level hierarchal data model, which is action-round-physiology. The algorithm is capable of balancing the energy expenditure of each round by calculating the energy consumed at high levels of intensity and recovering dynamic suggestions of training intensity and recovery on an individual basis depending on fatigue index, heart rate recovery curves, and muscle load thresholds and is based on continuous round-level energy consumption calculations, high-intensity scenario offsetting, and adaptive calibration of individual differences. The experimental findings indicate that the round-level energy consumption estimation model is able to represent the variations in the intensity of actions and offensive/defensive rhythm. Indicatively, with a high intensity round, P02 energy consumption was at 6.7 kcal round as compared to 6.0 kcal in a round of low intensity, which is a clear indication of exercise load variance. In the high-confrontation situations, the energy consumption correction mechanism also optimizes the calculation of the energy consumption. In an example, there is an increase in the energy consumption of round 1 of P02 of 6.7 kcal to 7.2 kcal after correction which is more representative of the real energy consumption of the continuous sprints and body contact.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"281 ","pages":"Pages 83-92"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141250","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Optimization Path of Modern Industrial System Based on Intelligent Algorithms","authors":"Dongwen Xie","doi":"10.1016/j.procs.2026.04.196","DOIUrl":"10.1016/j.procs.2026.04.196","url":null,"abstract":"<div><div>As the scale of modern industrial systems grows and it becomes more sophisticated, the conventional optimization approaches suffer seriously in the aspect of the granularity of resource allocation, the depth of collaboration of the industrial chain, and the agility of dynamic adjustment. They especially find it difficult to provide high-precision and real-time overall optimization, especially on multi-objective trade-offs, cross-link collaboration and the introduction of new industries. To solve this, this paper proposes the Artificial Fish Swarm Algorithm (AFSA) to develop an intelligent optimization model to the industrial systems to resolve some of the major issues including rough industrial node set-up, slow collaborative scheduling, and slow reaction to layout modification. The foraging, tailing, swarming, and random activities of AFSA are mapped to resource allocation, collaboration in the industrial chain, industrial cluster development as well as emerging industries introduction respectively. Adaptive step size, dynamic vision and multi-objective fitness functions are used to provide fine-grained parameter control and state updates. Moreover, the industrial network structure includes node flow, capacity limitation, energy use assessment, and environmental indicators that allow real-time tracking and adjusting the system aspects during the search process of the algorithm. As per the experimental findings, with resource allocation experiment, the overall efficiency index FFafter was 0.78, 0.73, 0.81, 0.76, and 0.84; in the industrial chain collaboration experiment, the collaboration index CIafter was 0.71, 0.74, 0.82, 0.68 and 0.79; with the emerging industry introduction experiment, the system revenue Rafter was 2.32 billion, 2.75 billion, and 2.51 billion, respectively. These statistics allow stating that AFSA is able to offer exact resource allocation plans to contemporary industrial systems, organize both upstream and downstream cooperation and be able to support stability of systems and increase their revenues once the new industries emerge.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"281 ","pages":"Pages 93-102"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141251","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Application of Computer Vision Big Data Algorithms in Accurate Detection of Corrosion Images of Boiler Water-cooled Wall Tubes","authors":"Zijian Zhang ,&nbsp;Zhaoqiang Luo ,&nbsp;Junping Zhong ,&nbsp;Lei Zheng ,&nbsp;Junhui Chai","doi":"10.1016/j.procs.2026.04.209","DOIUrl":"10.1016/j.procs.2026.04.209","url":null,"abstract":"<div><div>Boilers, as core and critical equipment in energy, chemical, and other fields, have water-cooled wall tubes that operate under harsh conditions of high temperature, high pressure, and corrosive media for extended periods. These tubes are prone to corrosion failure, directly threatening equipment safety and causing significant economic losses. Current corrosion detection of boiler water-cooled wall tubes largely relies on manual visual inspection or traditional non-destructive testing methods, which suffer from low efficiency, high subjectivity, insufficient accuracy, and difficulty in adapting to large-scale testing needs. To address these challenges, this paper proposes a precise corrosion image detection scheme based on computer vision big data algorithms, constructing a complete technical system encompassing \"data acquisition and preprocessing - feature extraction and fusion - corrosion identification and quantification.\" First, a multi-source image acquisition platform is built to obtain a large-scale dataset of water-cooled wall tube corrosion images, and data augmentation and denoising algorithms are used to improve data quality. Second, an improved convolutional neural network and attention mechanism are integrated to construct a feature extraction model, achieving accurate capture of deep semantic features and local detail features of the corrosion area. Finally, a multi-scale corrosion quantification assessment module is designed to complete corrosion level classification and damage degree quantification. Experimental results show that the proposed computer vision big data algorithm achieves the highest detection accuracy on both self-built datasets and publicly available industrial datasets, reaching 96.2% and 93.5% respectively. Compared with the ResNet-50 algorithm, the accuracy is improved by 10.4 percentage points and 12.3 percentage points respectively. It can effectively meet the needs of accurate detection of corrosion of boiler water-cooled wall tubes in industrial scenarios and provide reliable technical support for equipment life cycle management.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"281 ","pages":"Pages 207-215"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141340","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"An IoT-Based Intelligent Framework for Real-Time Dam Monitoring and Control","authors":"Imen Tounsi ,&nbsp;Mohamed Hadj Kacem","doi":"10.1016/j.procs.2026.04.032","DOIUrl":"10.1016/j.procs.2026.04.032","url":null,"abstract":"<div><div>Dams are critical infrastructures; their management and control are essential to ensure their safety, optimal utilization of water resources, and proactive maintenance. In this paper, we propose an innovative IoT-based intelligent framework for real-time dam monitoring and control. The system integrates a network of advanced sensors, edge devices, and cloud-based data analytics to enable real-time monitoring and predictive maintenance of dam infrastructures. The system utilizes ultrasonic water level sensors, temperature probes, and environmental sensors to collect critical data, which is then processed by microcontrollers equipped with integrated wireless connectivity. A key originality of our approach lies in the adoption of IoT design patterns, particularly the Delta Update and Incremental Update patterns, which significantly optimize data transmission, reduce bandwidth consumption, and improve energy efficiency. Data are securely transmitted via protocols to a cloud platform. Experimental results demonstrate the framework’s capability to enhance the accuracy, responsiveness, and efficiency of dam monitoring. Our approach not only addresses the limitations of conventional monitoring techniques but also provides a scalable, pattern-driven solution that supports data-driven decision-making for improved dam safety and operational performance.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"280 ","pages":"Pages 237-244"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141492","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Design of an AI-Driven Multimodal Personalized Learning Platform Based on Optical Learning Behavior Modeling","authors":"Bingjie Shen","doi":"10.1016/j.procs.2026.04.244","DOIUrl":"10.1016/j.procs.2026.04.244","url":null,"abstract":"<div><div>The currently used individualized college English learning systems are largely based on the history of interaction and audio signals, which makes the ability to assess oral skills susceptible to environmental noise, microphone fluctuation, and low clarity of articulation even with recent innovations in deep learning-based models of the learner featuring adaptive content delivery. In order to drive out these limitations, the paper re-engineers a personalized learning platform supported by AI-powered system by extending this to include an optical articulation modeling stream through traditional webcams. The combination of (i) learning behavior traces, (ii) speech audio and (iii) optical cues such as lip/jaw movements, facial landmarks and gaze/attention indicators are integrated into a closed-loop pipeline of multimodal sensing -&gt; representation learning -&gt;learner state inference -&gt;adaptive practice and feedback. The use of self-supervised and transformer-based fusion methods allows the system to represent audio-visual speech more robustly, even in noisy audio conditions, and light-weight optical pipelines (e.g. face mesh/landmark tracking) enable the real-time extraction of articulation features on consumer devices. Apart from enabling personalized vocabulary and reading practice, the improved platform offers formative feedback that centers on pronunciation, whereby phoneme-level acoustic confidence is matched with visual articulation consistency, and it measures engagement based on optical attention traits to promote intervention timing. Experiments on the learning data of the platform indicate high-quality predictions of the learner state (AUC values may be up to 0.835), and online A/B testing demonstrates the improvement of the rate of learning retention and the duration of the weekly study after the adaptive optimization. The proposed optics-enhanced design has a viable pathway to the enhanced oral-skill assessments and sensitive personalization of big college English learning.</div></div>","PeriodicalId":20465,"journal":{"name":"Procedia Computer Science","volume":"281 ","pages":"Pages 521-530"},"PeriodicalIF":0.0,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148141511","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}