YOLO-GC: A Lightweight Model for Remote Sensing Image Object Detection

Remote sensing images are crucial national strategic resources. However, existing high-performance object detection models for remote sensing images typically suffer from high computational complexity and large parameter sizes. To reduce computational demands while maintaining detection accuracy for multi-scale targets, this study proposes YOLO-GC, a lightweight object detection model for remote sensing images. First, the conventional Conv modules in the YOLOv8 backbone network are replaced with GhostConv, and the C2f module is enhanced into a C2f-GhostConv structure integrating Ghost operations, significantly reducing model complexity. Meanwhile, a CBAM attention layer is incorporated into the backbone network, effectively improving feature extraction capability without increasing computational overhead. Experimental results on the NWPU VHR-10 remote sensing object detection dataset demonstrate that YOLO-GC achieves a computational complexity of 21.4 GFLOPs, reducing the baseline YOLOv8s' complexity by 7.4 GFLOPs, while maintaining a mAP@50 of 0.922-only 0.011 lower than the baseline. YOLO-GC achieves a superior balance between accuracy and efficiency, significantly enhancing its potential for deployment on resource-constrained edge platforms.