International Journal of Innovative Computing and Applications (15 papers in press)

Regular Issues

• Wireless Inertial Motion Capture System and Pose Trajectory Fusion Algorithm for Animation and Film Design  
  by Shifeng Wang 
  Abstract: The film and animation industry increasingly demand real-time performance and accuracy. Traditional optical motion capture methods face limitations due to environmental lighting and equipment constraints, often failing to meet high production standards. To address this issue, this study proposes a motion capture system that integrates an error-state Kalman filter and improved adaptive damped least squares. The system fuses data from multiple inertial measurement unit sensors through the error-state Kalman filter and optimises pose trajectories with the improved adaptive damped least squares, significantly enhancing motion capture performance. Experimental results show that the proposed system achieves a pose estimation error converging to 1.5 degrees within 2.5 s, and the trajectories closely match the optical ground truth. During dynamic tasks, the system exhibited absolute orientation errors of 2.44
  Keywords: Motion capture system; Pose trajectory fusion algorithm; Inertial measurement unit; Kalman filter; Animation and film design.
  DOI: 10.1504/IJICA.2026.10078298
   
  
• A Framework for Indoor Visual Element Extraction Using an Improved Attention Generative Adversarial Network  
  by Miao Nie, Zhongping Sun 
  Abstract: Traditional generative adversarial networks (GAN) suffer from insufficient structural modelling and loss of details in indoor visual element extraction. This study proposes an improved attention GAN model framework that integrates spatial attention mechanisms and channel attention allocation mechanisms. This framework guides the generator to continuously optimise the feature expression process through the supervised feedback mechanism of the discriminator, thereby improving the accuracy and completeness of visual element extraction, ensuring feature recognition and reconstruction in complex image scenes. The experimental outcomes showed that the algorithms element extraction accuracy was 93.5%, the edge segmentation accuracy was 91.2%, and it exhibited stronger robustness in interference environments. The indoor element extraction model achieved an element extraction accuracy of over 90.3% on different datasets, demonstrating good convergence and generalisation ability. The structural similarity of the extracted element feature images reached 0.94, with a PSNR of 37.4 dB. The above results indicate that the improved attention GAN model can effectively identify various elements in complex indoor scenes and maintain a high degree of detail restoration in indoor visual element extraction. This study provides more reliable technical support for indoor environment understanding and intelligent design.
  Keywords: Indoor visual element extraction; Generate adversarial networks; Attention mechanism; Spatial attention; Channel attention.
  DOI: 10.1504/IJICA.2026.10078660
   
  
• Intelligent inspection and monitoring of minor defects in complex slopes using unmanned aerial vehicles based on improved YOLOv11-PEW architecture  
  by Yong Lu, Li Tian, Sheng Yuan 
  Abstract: To address the issues of missed detection of small targets, interference from complex backgrounds, and blurred boundaries in visual inspection of geological disasters, this paper proposes the YOLOv11-PEW model: it introduces a P2 microscopic inspection head to retain shallow features to reduce texture loss, embeds an efficient multi-scale attention (EMA) module to calibrate feature weights across space to suppress background noise, and uses the Wise-IoU (WIoU) loss function (LF) to optimize boundary regression accuracy. Experimental findings demonstrate that YOLOv11-PEW achieves a mAP@0.5 of 90.1% on the self-built slope dataset, a 5.8% improvement over the baseline (YOLOv11n). Furthermore, while ensuring compliance with the real-time inference standard for embedded edge devices (>30 FPS), it significantly improves the detection accuracy of extremely small targets (Area <32 x 32) by 14.6%. Visualisation analysis (Grad-CAM) further confirms that the model can accurately focus on the disease itself from diffuse background noise, demonstrating excellent robustness.
  Keywords: highway slope protection; detection of minor defects; YOLOv11; attention mechanism; unmanned aerial vehicle remote sensing.
  DOI: 10.1504/IJICA.2026.10078678
   
  
• Deep Learning-Based Traffic Flow Multi-Source Heterogeneous Information Fusion and Emergency Management Response Mechanism  
  by Sirui Chen 
  Abstract: Accurate perception of fabric deformation is an important prerequisite for applying wearable sensor networks to intelligent clothing design. However, purely data-driven methods lack physical consistency, making it difficult for the wearable sensor network to perceive and predict fabric deformation accurately. To address these issues, this paper proposes a physical-informed joint simulation-recognition sensor framework for fabric. The core idea of achieving a precise perception of fabric deformation by calibrating fabric physical properties in the sensing data, focusing on key parts prone to deformation, and optimising model performance through closed-loop feedback is realised. Experimental results show that the proposed method achieves 0.941 in prediction accuracy for the fabric deformation degree and 0.86 in recognition accuracy for clothing fit, which are 5.5% and 8.9% higher than those of the optimal method. This research presents a new model for applying wearable sensor networks to the perception of fabric deformation.
  Keywords: Traffic Flow; Multi-source Heterogeneous Information Fusion; Emergency Management Response; Multi-head Attention Mechanism; Graph Convolutional Network.
  DOI: 10.1504/IJICA.2026.10079541
   
  
• A cooperative optimization method for coverage and scheduling of traffic WSN relay nodes considering carbon emission cost constraints  
  by Xiaxiao Tian 
  Abstract: To address the high energy consumption and carbon emissions of relay nodes in wireless sensor networks for traffic sensing, this paper proposes a collaborative optimisation method for relay coverage and scheduling under carbon emission cost constraints. This method constructs a joint scoring mechanism of coverage gain and carbon emission penalty, embeds carbon constraints during the node activation stage, and couples the coverage optimisation results with the dynamic duty cycle adjustment closed loop through carbon budget allocation, forming a collaborative decision-making link for coverage selection, carbon quota decomposition, and scheduling execution. Experimental results show that the proposed method converges within 40 iterations, achieving a coverage rate of 92%, which is superior to the multi-objective differential evolution algorithm (83%) and the particle swarm optimisation algorithm (75%). This method enables efficient and low-carbon operation of green transportation sensing systems, providing a feasible solution for their efficient and low-carbon operation.
  Keywords: Traffic wireless sensor network; Relay node; Carbon emission constraints; Coverage optimization; Scheduling coordination.
  DOI: 10.1504/IJICA.2026.10079636
   
  
• Hierarchical Asynchronous Federated Learning Method for Blockchain-enabled Vehicle Networking Integrating Spatio-temporal Trajectory Characteristics  
  by Li-ting Zhang, Yunfei Wang, Hou Fu Zhang, Ziyang Chen, Zeng Sheng 
  Abstract: As the Internet of Vehicles (IoV) transitions into a data-driven ecosystem, it encounters critical challenges regarding privacy preservation and resilience against model poisoning attacks. Conventional Federated Learning (FL) is constrained by the highly dynamic topology of vehicular networks and lacks intrinsic trust mechanisms to mitigate stealthy poisoning attacks. To bridge these gaps, we propose B-HAFL, a Blockchain-enabled Hierarchical Asynchronous Federated Learning framework. First, to address network efficiency, we construct a vehicle-to-cluster-to-cloud architecture that utilizes edge-side pre-aggregation, effectively reducing communication overhead by approximately 70%. Second, to fortify system security, we design a physical-logical fused multidimensional dynamic reputation model. This model not only incorporates FastDTW-based trajectory analysis to verify physical behaviour but also integrates a time decay factor to dynamically evaluate the timeliness and reliability of node contributions. Based on this, a reputation-driven differentiated weighted aggregation (RDWA) algorithm and a smart contract-based incentive mechanism are implemented. Extensive simulations on the UNSW-NB15 dataset demonstrate the frameworks robustness. Specifically, in a hostile environment with a 30% poisoning attack rate, B-HAFL maintains a classification accuracy of 92%. Furthermore, by incorporating the time decay factor, the models accuracy under the same 30% attack scenario is significantly improved to 95.8%, effectively preventing model collapse compared to baseline methods.
  Keywords: Internet of Vehicles (IoV); blockchain; asynchronous federated learning; secure data sharing; dynamic reputation; physical-logical fusion.
  DOI: 10.1504/IJICA.2027.10079755
   
  
• DHT Load Balancing Algorithm for Distributed File Systems  
  by Dejun Miao, Rongyan Xu, Jiusong Chen, Yan Wu 
  Abstract: To address load imbalance and node congestion in distributed file systems deployed in intelligent aviation cabins, a distributed hash table-based load balancing algorithm named MTTL-K is proposed. The algorithm integrates predictive data migration and adaptive threshold control into a decentralised storage architecture to improve balancing efficiency and stability. A dynamic migration mechanism is designed to optimise target-node selection and reduce unnecessary migrations. Experimental results show that the proposed algorithm reduces the load standard deviation to 3.69% after multiple migration rounds and decreases total migrations to 62% of the initial level. Compared with the conventional TTL-K algorithm, MTTL-K reduces the imbalance degree to approximately 0.5 within two minutes while maintaining low migration overhead. These results demonstrate that the proposed method improves load distribution, resource utilisation, and operational stability in distributed file systems.
  Keywords: MTTL-K algorithm; Distributed Hash Table; Dynamic Load Balancing; Predictive Migration; Aviation DFS System.
  DOI: 10.1504/IJICA.2027.10079885
   
  
• A Two-Stage Hybrid Computing Framework Integrating Enhanced 1D-CNN and Graph Clustering for Distribution Network Topology  
  by Xun Xu, Jiacheng Li, Chang Ye, Xuan Cai 
  Abstract: This paper proposes an innovative neuro-computing paradigm to overcome extreme class imbalance and high-noise temporal misalignment in real-world time-series data. We introduce a two-stage hybrid machine learning framework, successfully applied to smart grid topology identification. The first stage develops an enhanced 1D-CNN-Transformer combining Squeeze-and-Excitation attention and Focal Loss to dynamically reshape decision boundaries for long-tail classification. The second stage employs unsupervised Graph Laplacian Spectral Clustering, constrained by Dynamic Time Warping (DTW), to solve synchronization errors in extremely small samples. Experimental results demonstrate that our neuro-computing model achieves 96.8% supervised classification accuracy under extreme imbalance. Furthermore, the unsupervised graph clustering attains 97.22% overall accuracy, maintaining robust performance even with 15% missing data. Ultimately, this framework accurately corrects system anomalies (reducing line loss from 9.2% to 6.8%), proving its exceptional reliability as a novel computing solution for complex industrial applications.
  Keywords: neuro-computing; machine learning paradigms; 1D-CNN-Transformer; graph Laplacian clustering; extremely imbalanced classification; time-series sensing; smart grid.
  DOI: 10.1504/IJICA.2027.10080419
   
  
• An Intelligent Framework for Forecasting Cardiovascular Disease using Bio-inspired Optimization and ML  
  by Aditya Kumar, Abhijeet Kumar, Ashish Mishra 
  Abstract: Globally, heart disease is still the world's top cause of death Hence, early and accurate diagnosis is crucial to enhancing patient outcomes and reducing death rates The paper offers a machine learning-based framework for heart disease prediction utilising the UCI Cleveland da-taset in order to allay this worry To improve prediction reliability, the method combines sophisticated boosting models like LightGBM and XGBoost with the Jellyfish Op-timisation Algorithm for accurate feature selection Prior to choosing the most important characteristics, first prepro-cessing addresses missing data and normalises attrib-utes The dataset is split 80:20 between training and test-ing sets RandomizedSearchCV is used to optimise LightGBM, and classification criteria are modified to im-prove accuracy Accuracy, precision, recall, sensitivity, specificity, F1 score, and AUC are used to gauge perfor-mance According to the results, LightGBM with JFO out-performs XGBoost in terms of accuracy (91 25%,86 67%) and AUC (96 77%, 95 14%), indicating an effective approach.
  Keywords: Machine Learning; JFO; Heart Disease; XGBoost; LightGBM; RandomizedSearchCV.
  DOI: 10.1504/IJICA.2027.10080542
   
  
• Image Element Generation and Application in Visual Messaging Design Based on an Improved CycleGAN  
  by Lanlan Zheng 
  Abstract: In visual messaging design, image element generation impacts style consistency and production efficiency. However, current methods rely on designers experience and traditional tools, leading to low standardization and limited efficiency. This paper proposes a Cycle Consistent Generative Adversarial Network (CycleGAN)-based model for generating and applying image elements. Leveraging strong feature extraction and mapping capabilities, the model streamlines the generationapplication process. Experiments show a loss value of 0.02significantly lower than comparison modelsrequiring fewer iterations. It achieves 0.94 accuracy and recall, with an area under the curve of 0.98, all outperforming baselines. The model effectively solves low standardization and poor adaptability issues in image element generation, offering an efficient, reusable, and standardized approach for visual communication design.
  Keywords: Cycle GAN; Visual communication design; DO-Conv; Image element generation.
  DOI: 10.1504/IJICA.2027.10081023
   
  
• A Multi-Criteria Optimisation Framework for Expertise-Driven Bug Assignment with Dynamic Workload Balancing  
  by Saurabh Saxena, Amarjeet Prajapati 
  Abstract: Efficient bug assignment remains a critical challenge in large-scale software development due to conflicting objectives such as expertise matching, priority handling, workload balance, and reassignment. Traditional approaches either focus solely on expertise, leading to developer overload, or emphasize load balancing while neglecting priority sensitivity. This paper proposes a priority-aware multi objective framework that integrates developer ranking, dynamic workload evaluation, and controlled task redistribution into a unified decision model. The assignment process is formulated as a multi objective optimization problem that maximizes priority-weighted expertise alignment while minimizing workload variance, expected resolution time, and unnecessary reassignment. Extensive simulated experiments demonstrate that the proposed approach achieves 91.4% high-priority resolution accuracy, reduces workload variance by up to 85% compared to expertise-only assignment, lowers average resolution time, and significantly limits task tossing. The results indicate that the proposed framework effectively balances efficiency and system stability, making it suitable for large-scale and dynamic software development environments.
  Keywords: Bug Assignment; Software Engineering; Multi-objective criteria; Task Optimization; Workload Balancing.
  DOI: 10.1504/IJICA.2027.10081191
   
  
• Optimisation of Intelligent Image Matting Technology for Film and Videos Based on Transformer and KD  
  by Xiao Wang, Yumeng Wu 
  Abstract: To tackle edge discontinuity and semantic misclassification in video segmentation, this study proposes ST-ViMat, a dual-stream collaborative model integrating Swin Transformer with similarity-preserving knowledge distillation (SPKD) and a squeeze-and-excite attention mechanism for enhanced multi-scale edge representation. Experiments show ST-ViMat achieves a SAD score of 50.2?2.5 and temporal distortion error of 0.020?0.002, outperforming mainstream methods. Ablation studies confirm SPKD reduces SAD by 10.3%, and the SE-RCF edge module further improves it by 6.3%. The model maintains robust performance across resolutions from 1080p to 480p, with only a 7.0% SAD decay. Subjective evaluation yields an edge smoothness score of 9.30?0.64. With just 25 million parameters, ST-ViMat leverages cross-window feature alignment and channel attention to ensure edge continuity and semantic consistency in dynamic scenes, offering an efficient, high-quality solution for film-grade video segmentation.
  Keywords: Film and video processing; Swin Transformer; SPKD; Image features; Edge segmentation.
  DOI: 10.1504/IJICA.2027.10081197
   
  
• Optimization Model for Fake News Detection Based on Large Model and LoRA Retrieval-Augmented Generation  
  by Liping Shen, Junyue He, Lu Ban, Liheng Zhang 
  Abstract: Challenges are encountered by traditional detection methods when dealing with fake news. This is because the methods lack competence in semantic understanding, anti-interference, and knowledge updates, while the spread of fake news inflicts serious social damage. Therefore, this study proposes an optimization model for false news detection based on large models and low rank adaptive fine-tuning, and introduces retrieval enhanced generation and high-quality vector embedding models to achieve multi-dimensional evidence fusion and efficient classification. This study uses prompt engineering and large model optimization to input data, and then uses low rank adaptation for efficient parameter tuning. Regarding resource efficiency, the video memory consumption of the research model was limited to around 10 GB, and its throughput was sustained at 138-141 strips/s. High accuracy was guaranteed, and efficiency and scalability were simultaneously balanced. The research findings fully verify the methods validity and utility in fake news detection.
  Keywords: False news detection; Large model; Low-rank adaptation; Retrieval-augmented generation; High-quality vector embedding model.
  DOI: 10.1504/IJICA.2027.10081551
   
  
• M-DSTAR: An Innovative Spatiotemporal Computing Framework for Efficient Video Action Recognition in Sports Applications  
  by Wei Shen, Jie Wu, Xuekai Qi, Li Liu 
  Abstract: Video action recognition for intelligent sports requires balancing computational efficiency and accuracy. We propose M-DSTAR, a computing framework tailored for practical sports video analysis. To address computational redundancy and variable-speed motion challenges, we introduce a decoupled spatiotemporal attention (DSTA) module and a speed-adaptive dynamic window (SADW) mechanism. DSTA computes spatial and temporal features independently to optimise representation learning. Concurrently, SADW dynamically adjusts the temporal receptive field based on real-time motion intensity. Applied to the UCF101 and HMDB51 datasets using only ImageNet-1K pre-training, M-DSTAR achieves an accuracy of 94.3% and 72.1% under dense offline evaluation. Crucially, under a practical online single-view protocol with micro-latency buffering for real-time streaming, the model maintains an accuracy of 91.9% at 76 FPS. Furthermore, the architecture reduces floating-point operations (FLOPs) by 46% compared to heavy 3D/2D hybrid models, maintaining a favourable efficiency-accuracy trade-off against lightweight CNN baselines. This framework demonstrates practical viability for complex motion recognition on standard computing platforms.
  Keywords: innovative computing; video action recognition; spatiotemporal attention; dynamic receptive field; edge deployment; intelligent sports applications.
  DOI: 10.1504/IJICA.2027.10081741
   
  
• Multi-dimensional resource optimisation and network slicing for 6G network based on reinforcement learning  
  by Zhuoran Wang, Chi Zhang 
  Abstract: The sixth generation (6G) of mobile communication technology has emerged as a leading technology, ready to connect everything in the future. 6G networks are expected to enable low-latency communications, massive terminal connectivity, and previously unheard-of mobile broadband speeds. This is due to the continuous advancements in science and communication technology. This paper proposes a novel algorithm for joint optimisation of network slicing and task offloading in 6G network. The main idea is to jointly offload multiple mobile edge computing (MECs) devices and minimise the energy consumption and execution delay under limited resources based on reinforcement learning. Simulation results show that the proposed algorithm has superior performance as compared with existing algorithms.
  Keywords: network slicing; 6G network; reinforcement learning; mobile edge computing; MECs.
  DOI: 10.1504/IJICA.2027.10081891