A Modified YOLOv5 Architecture for Aircraft Detection in Remote Sensing Images

The primary challenge in detecting aircraft in remote sensing images arises from various factors, such as diverse aircraft categories, small objects, and intricate backgrounds. The single-stage algorithms, such as YOLOv5, are excel in detection and offer real-time performance with a simple architecture, but they often face issues with lower accuracy. To tackle this, a new model based on YOLOv5 is proposed, that integrates a window-based self-attention mechanism through the Swin Transformer v2 block into the C3 module in the Backbone. This aims to improve the utilization of both global and local information. Additionally, a bi-directional feature pyramid network (BiFPN) structure is integrated into the Neck to enhance feature fusion at various scales. Finally, a dedicated small object detection head is introduced in the Prediction part to enhance the detection accuracy of small aircraft. Validation on the Military Aircraft Recognition dataset (MAR20) against YOLOv5, YOLOv7, and YOLOv8 shows superior results, with our modified YOLOv5 model achieving a mean average precision (mAP) of 69.8% and a mean average precision at an intersection over union threshold of 0.5 (mAP0.5) of 92.2%.