A Direction-Sensitive Method for Real-Time Small Object Detection in Remote Sensing

Object detection in remote sensing imagery is a fundamental and crucial task within the remote sensing domain. In this field, mainstream methods can be divided into two-stage and one-stage approaches, among which the one-stage methods achieve an excellent balance balance between accuracy and efficiency. Representative of one-stage methods is the YOLO series, among which YOLOv10 stands out for its remarkable accuracy-efficiency trade-off. However, small object detection in remote sensing images is a challenge for state-of-the-art (SOTA) methods. The reason lies in the fact that small objects often have a large aspect ratio, appeared significant variations in its characteristics across different orientations, distributed against complex backgrounds. SOTA methods, which employ square convolutions for feature extraction, dopt an undifferentiated approach to feature extraction across all orientations, often retaining excessive background information during feature extraction, leading to suboptimal detection performance. In this context, an orientational-aware foreground-aware small object detection network based on YOLO is proposed to achieve small object detection in complex backgrounds. OSFA-YOLO employs a orientational feature enhancement module to enhance the directional and shape features of small objects from multiple orientations, improving the sensitivity of network to the intrinsic features of the small objects. Additionally, it incorporates a feature fusion module based on attention mechanism, which integrates information of the small objects as guiding cues with other features, enabling suppressing background interference while preserving the extracted small object features.