Recently, the research paper titled "Adaptive Feature Fusion and Task-Dynamic Alignment for Real-Time Object Detection on Edge Devices" by Professor Zhao Jia's team from CIT's School of Computer Technology and Engineering was published in Expert Systems with Applications, a top-tier international journal in the field of artificial intelligence (DOI: https://doi.org/10.1016/j.eswa.2025.128341). The first author of the paper is Associate Professor Ding Yan, with Professor Zhao Jia serving as the corresponding author. Expert Systems with Applications is a leading international journal in AI, operations research, and intelligent systems, recognized as a CAS Q1 TOP journal with an SCI impact factor of 7.5.
Due to the limited computing power, high real-time requirements, and complex application scenarios of resource-constrained edge devices, traditional object detection methods face the dilemma of balancing accuracy and efficiency.
To address this, the team proposed EdgeSight, a highly efficient object-detection architecture for edge-computing environments, which innovatively integrates lightweight feature extraction, task-dynamic alignment, and intelligent model compression to address the issue that existing methods fail to meet both high-accuracy and low-latency requirements simultaneously. By ingeniously combining RepViTBlock modules, deformable convolutional detection heads, and knowledge-distillation mechanisms, EdgeSight balances detection accuracy and computational efficiency. Experiments results show it performs well on multiple benchmark datasets, and supports edge computing applications in smart video analysis, industrial inspection, and environmental monitoring, and other fields.
First Review by Zhan Xia
Second Review by Zhang Jian
Final Approval by Su Xiaoping
