Improved YOLOv5 in Remote Sensing Slender and Rotating Target Detection

Although the deep-learning based object detecting methods for remote sensing imagery have achieved very high accuracy on most of the targets, the slender and rotating targets in remote sensing images, such as bridges and harbors, are exceptional. The problem is that the predefined anchor boxes usually cannot well cover these targets. Thus, in this study an improved YOLOv5 model is proposed to better detect the slender rotating targets in remote sensing imagery. Firstly, the size of input images is doubled to maintain the integrity of the target. Secondly, the deformable convolution (DCN) is applied in the backbone net-work during feature extraction to increase the coverage of the feature and decrease the interruption by the background while extracting the slender rotating targets. Finally, besides length and width of a bounding box, the aspect ratio of the bounding box is also added to the loss function to emphasize the im-portance of the uniqueness of the slender rotating targets and to improve the prediction accuracy subse-quently. The three modifications have been applied to the YOLOv5 model respectively and together on DOTA dataset. The results show that the proposed method can better detect slender and rotating targets in remote sensing images and the accuracy for bridge and harbor has been improved to 52.1% and 75.16% respectively.