Small Object Detection in Aerial Images based on Cross-Attention and Semantic Denoising

Small object detection in aerial images remains a challenging task due to scale variation, dense distribution, and background noise. To address these limitationse, this paper proposes a lightweight improvement to YOLOv8 by introducing two modules: Cross Attention, which replaces naive concatenation in the Feature Pyramid Network (FPN) to enable selective cross-scale interaction, and Semantic Gate eXchange (SGX), which adaptively suppresses background noise after feature fusion using global semantic pooling and local noise estimation. Experiments on the AI-TOD dataset show that our method consistently improves performance over the YOLOv8-nano baseline, achieving a +10% gain in mAP@50 and a +3% gain in mAP@50-95, while maintaining efficiency suitable for deployment on edge devices. These results demonstrate that the proposed framework provides a practical balance between accuracy and efficiency for aerial object detection.