Publications
Sort:
Open Access Issue
A semantic-aware edge-cloud collaborative inference framework for small object detection
Journal of Northwest University (Natural Science Edition) 2026, 56(5): 1150-1160
Published: 25 October 2026
Abstract PDF (921 KB) Collect
Downloads:0

Small object detection in UAV inspection and intelligent surveillance scenarios typically relies on high-resolution inputs to capture fine-grained features. However, edge devices are constrained by limited computational resources and bandwidth, making it difficult to maintain detection accuracy while ensuring real-time performance. Existing edge-cloud collaborative inference methods primarily optimize computational overhead and communication cost, while overlooking the impact of feature compression on the semantic information of small objects, which often leads to the loss of critical features and degraded detection performance. To address these challenges, this paper proposes a semantic-aware edge-cloud collaborative inference framework. Specifically, a semantic-aware cost model is constructed to explicitly incorporate the semantic loss caused by feature compression into the scheduling decision, enabling joint optimization of latency and detection accuracy. Meanwhile, a content-adaptive time-balance feedback control mechanism (TBFC) is designed to dynamically adjust the partition point and compression strategy. In addition, a CPU-GPU heterogeneous collaborative execution pipeline is developed to improve resource utilization and reduce end-to-end latency. Extensive experiments on the VisDrone-DET and MS COCO datasets demonstrate that the proposed method achieves average end-to-end latencies of 132.5 ms and 128.6 ms, with mAP@0.5 of 24.1% and 52.2%, respectively. Compared with existing baseline methods, the proposed framework effectively reduces inference latency while maintaining detection accuracy under dynamic network conditions, demonstrating strong stability and practical applicability.

Total 1