UAV image detection algorithm based on improved YOLOv5

Abstract

The target detection algorithm based on deep learning usually has a good effect on traditional target detection, but has a low detection accuracy on small targets. To resolve the problem, this paper proposes an improved YOLOv5 algorithm for small object detection by research on the images captured by UAVs. First, aiming at the problem of high sampling frequency and large image receptive field, upsampling was added to further expand the feature map. Then, to solve the problem of insufficient semantic information of shallow features, feature fusion method was used and one 160*160 output detection layer was added. Finally, since the above steps would increase the amount of computation, the Mobilenet-V2lightweight network was added to improve the detection speed. The VisDrone UAV image data set was used for training and verification and the self-built data set was used for testing. The experiment results show that that improved YOLOv5 algorithm can effectively detect and identify small objects. Compared with traditional detection methods, the small objects detection accuracy and speed is also improved.