Object Detection Based on Feature Balance Pyramid in UAV Imagery

Compared with images taken from the ground perspective, small objects account for a large proportion of aerial UAV images and the image perspective changes greatly, which affects the target detection effect of aerial UAV images. In this paper, the Yolov5 algorithm is improved to adapt to UAV object detection. Given a large number of small objects in aerial images, the feature balance pyramid structure is added to improve the loss of low-level features and improve the detection effect of the small object. In the feature balance pyramid, Pixel un-Shuffle is used to adjust the scale of the feature, which preserves the low-level feature information and reduces the computational cost. The cross self-attention module is proposed to improve the balanced feature map and improve the positioning accuracy of the small object. The Angle of view of aerial images varies greatly. In this paper, the deformable convolutional network is added to the backbone network of Yolov5 to enhance the feature extraction capability of the model for multi-view objects. Experimental results show that on the visdrone data set, the improved algorithm improves the average accuracy (mAP) by 1.4 percentage points compared with the original algorithm.