Aircraft Rotated Boxes Detection Method Based on YOLOv5

Aiming at the problem of arbitrary alignment direction of aircraft targets in remote sensing images, a rotation frame detection method for aircraft targets is proposed based on YOLOv5. First, the circular smoothing label is introduced in the network classification regression to process the target angle information to avoid the abrupt change of angle value; then, the convolution operation in the backbone network is replaced by involution to reduce the number of network parameters; finally, Coordinate Attention is introduced after the Neck part to remove the background information from the features. The improved network is trained and tested on the aircraft target in the DOTA dataset, and the AP value is 90.68%, which is higher than other rotating frame detection algorithms, verifying the effectiveness of the network in this paper.