FR-YOLOv7: feature enhanced YOLOv7 for rotated small object detection in aerial images

Abstract

Abstract For the problems of weak object feature expression and variable object orientation in aerial image object detection, this paper proposed a feature enhanced YOLOv7 for rotated small object detection in aerial images. Firstly, for the problem of feature loss in the feature extraction stage, the feature enhanced spatial pyramid pooling and cross stage partial connections module was proposed, which effectively boost the feature expression of small object. Secondly, an attention guided max-pooling module was constructed to address the problem of feature loss. Then, the rotated object detection head was introduced to solve the problem of false negatives caused by variable object angles and dense object distribution. Finally, a multi-scale loss function was proposed for improving the detection effects of rotated small objects detection in aerial images. Extensive experiments were conduct on the public datasets of DOTA and University of the Chinese Academy of Sciences-AOD, with the help of the proposed method, we can achieve the detection accuracy that the mean average precision are 79.7% and 98.9%, respectively. Experimental results demonstrate that the proposed method has a significant improvement on the detection of small targets in aerial images.