Research on lightweight object detection in optical remote sensing images

Remote sensing object detection technology is playing an increasingly important role in various military and civilian applications. To achieve low complexity and high detection accuracy, this paper proposes a high-precision and low-complexity remote sensing object detection method based on the YOLO algorithm. To enhance model efficiency, a lightweight backbone network, CR module, is designed to fully exploit the correlation and redundancy among feature maps. Additionally, an improved spatial pyramid pooling structure and depthwise separable convolutions are introduced to reduce model parameters and computational costs. Furthermore, to improve detection accuracy, a cross-stage local multi-scale densely connected feature pyramid network is constructed to optimize feature extraction. Experimental results on the NWPU VHR-10 dataset demonstrate that, compared with the baseline network, CRF-YOLO reduces parameter size to 27.9% of the baseline, improves detection speed by 9.25%, and enhances detection accuracy by 3.6%, outperforming state-of-the-art models in overall performance.