Small object detection in aerial view based on improved YoloV3 neural network

In application scenarios such as UAV inspection, deep learning-based object detection methods are increasingly used to improve the automation of line inspection. In the aerial view scene, the drone is usually fly at a high altitude from the ground, so the proportion of the object in the image is relatively small. When the YoloV3 network identifies small objects, the detection result would not be good because there is less information in the 8x downsampling feature map. In this paper, base on the LaSOT data set, the YoloV3 network has been modified by adjusting the values of anchors and establishing the 4x downsampling prediction layer to enhance the detection effect of small objects. Compared with the original YoloV3 network, the improved YoloV3 network has a certain improvement in convergence ability and detection accuracy compared to the original YoloV3 network.