An Improvement Method of YOLO Framework for Airplane Target Detection in Remote Sensing Image

An improvement method based on the YOLO framework for detecting densely packed airplane targets in visible light remote sensing images is proposed in this paper. The improvement focuses on three key areas: data augmentation, attention mechanisms, and prediction module optimization. First, the Copy-Paste technique is introduced to increase both the number of targets and the proportion of target information in the images. Second, the Normalization-Based Attention Module (NAM) is employed to strengthen channel and spatial attention, thereby highlighting local and salient features. Lastly, CIoU_Loss, in combination with the DIoU_NMS strategy, is applied to optimize non-maximum suppression, enhancing the model's ability to detect occluded targets in dense environments and speeding up the convergence process. Experimental results on the DOTAv1, NWPU VHR-10, RSOD, and custom datasets show that the proposed method significantly reduces the false positive rates compared to the common YOLOv5 model, and excel at detecting small and occluded targets.