Small Target Detection for Unmanned Aerial Vehicle Images Based on YOLOv5l

Abstract

Object detection on the unmanned aerial vehicle(UAV)image is a popular and challengeable research orientation. UAV images consist of a large number of targets, complex backgrounds, and affluent small targets. So as to accurately detect small targets on UAV images, we put forward a small target detection model with a type of attention mechanism based on YOLOv5. Firstly, the ECA attention mechanism module[1] is added after the structure block of CSPdarknet[2]in the backbone of YOLOv5 to focus on the region of interest. Then we improve the structure of YOLOV5 by adding an extra prediction head which is useful to detect different-scale objects. We also substitute the sampling method for transpose convolution[3] to keep more features. To reduce the parameters and expedite the training speed, we use GhostConv[5]to reduce parameters and train time. A mass of experiments on the dataset VisDrone21 [6] shows that the improved YOLOv5l has good performance on UAV images. As a benchmark model, YOLOv5l is more suitable for UAV small target detection due to its network structure. Under the YOLOv5l benchmark, the mAP50 of our model increased by 10.3% and the mAP50:95 increased by 7.4 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> . The mAP50 of the model trained at the input resolution of 1536*1536 can reach 62.8%.