BFR-RetinaNet: An Improved RetinaNet Model for Vehicle Detection in Aerial Images

Vehicle detection in aerial images has been applied in many fields and attracted more and more scholars' attention. In the task, the objects are multidirectional and arranged densely, the background information is complex, and the scale of the object is different. To achieve better detection performance, an improved detection model BFR-RetinaNet is proposed, which is based on the single-stage object detection model RetinaNet. BFR-RetinaNet optimizes vehicle positioning by adding a directional anchor box regression branch. Simultaneously, the model introduces a balanced feature pyramid structure to enhance the extraction of features and reduce the interference of complex backgrounds. The experimental results on the aerial dataset show that the precision, recall, and average precision of the proposed model have been improved to varying degrees. It achieves 86.2% Precision, 98.4% Recall, and 90.8% Average Precision (AP), which is 7.96, 0.45, and 0.58 points higher than R-RetinaNet.