Abstract
Abstract
Unmanned Aerial Vehicle (UAV) image object detection has broad applications. However, UAV images feature dense small objects and severe scale variations, which hinder traditional algorithms. Moreover, most existing methods prioritize accuracy while neglecting computational cost, limiting deployment on edge devices. To address these issues, this study proposes MSC-YOLO, a multi-module optimized UAV image object detection algorithm based on YOLOv11. First, a Multi-branch Structure-based Downsampling (MSD) module is designed to significantly reduce the parameter count and computational overhead while mitigating feature information loss. Second, a Shallow-based Bidirectional Feature Pyramid Network (SBiFPN) is introduced into the neck network to integrate shallow features, enabling it to capture finer details and thereby improving the detection accuracy for small objects. Third, a Context-aware Spatial Pyramid Pooling (CSPP) structure is proposed, which selects context information tailored to the receptive field as feature weights according to pooling kernel size, reducing the loss of target details during feature extraction. Finally, to further enhance the lightweight performance, DepGraph channel pruning is utilized to compress the improved algorithm, forming PMSC-YOLO. On VisDrone2019, PMSC-YOLO reduces Params and FLOPs by 60% and 47% compared to the baseline, without compromising detection accuracy. To verify generalization, we evaluate it on VisDrone2019 and RSOD against state-of-the-art and UAV-specific detectors. PMSC-YOLO achieves a favorable balance between detection accuracy and computational efficiency, well-suited for practical UAV deployment.