A remote sensing image object detection method ABS-YOLO based on improved YOLOv11

The problem of object detection in Earth remote sensing images is studied, which is important for agricultural monitoring, urban planning, early warning of natural disasters, etc. Due to the different sizes of objects, complex background, and dense distribution of small objects in remote sensing images, problems such as high percentage of missed objects and insufficient accuracy of their coordinates often arise. In this regard, an improved method for YOLOv11, ABS-YOLO, is proposed, which significantly improves the performance of object detection by integrating Averaged Convolution (AConv), Bidirectional Weighted Feature Pyramid (BiFPN), and Swin Transformer attention mechanism. Experimental results show that, compared with YOLOv11, the proposed object detection method ABS-YOLO with AConv, BiFPN, and Swin Transformer achieves 3.9 % increase in mAP50 estimations and 2.6 % increase in mAP50-95 on the NWPU VHR-10 dataset with significant improvement in precision and recall rates. This method allows achieving a balance between efficiency and accuracy of remote sensing object detection due to the proposed improvements.