SCAF-YOLO: Multi-Scale Feature Fusion for Small Object Detection in Remote Sensing Images

Small object detection in remote sensing images is challenging due to their limited size, indistinct features, and high background similarity. To address these challenges, we propose a method called Spatial Context-Aware Feature Fusion YOLO (SCAF-YOLO). The main contributions of SCAF-YOLO are as follows: First, the Spatial Context Aware Module (SCAM) is adopted to enhance the global context representation of small objects in images. Second, the Adaptive Spatial Feature Fusion (ASFF) detection head is incorporated to replace the original detection structure of YOLOv5, improving multi-scale feature fusion and selection capabilities. Finally, a Multi-Scale Non-Local Attention (MS-GLCA) is adopted. It combines multi-scale convolutions and channel attention to enhance feature representation via weighted fusion. This leads to more efficient feature weighting. The effectiveness of the proposed method is validated on two public remote sensing datasets, USOD and DOTA. Results show that the SCAF-YOLO model outperforms other state-of-the-art detectors in multi-object detection accuracy.