YOLO-FAD: enhancing small-object detection in high-resolution remote sensing images

Object detection technology serves as a crucial tool in remote sensing imagery analysis. However, the significant variations in object scales and complex backgrounds in remote sensing images pose substantial challenges, particularly for small object detection. To address these challenges, this paper proposes an improved network architecture based on YOLOv8n (You Only Look Once). First, the FAB (Feature Augmentation Block) module is introduced into the shallow and intermediate layers of the backbone network, enhancing feature representation through multi-scale convolution and residual connections to improve sensitivity to small objects. Second, the Dattention (Deformable Attention) module is incorporated into the deeper layers, leveraging deformable attention mechanisms to effectively focus on object regions while mitigating interference from complex backgrounds. Lastly, the original convolution modules in the Neck are replaced with the AKConv (Adaptive Kernel Convolution) module, which dynamically adjusts kernel sizes to enhance adaptability to multi-scale objects and improve feature extraction. Experimental results on the NWPU VHR-10 dataset demonstrate that the proposed model achieves a 2.5% improvement in mAP compared to the original YOLOv8n model, while reducing the parameter count to 2.6M, achieving lightweight optimization alongside enhanced detection accuracy.