Small Object Detection Method based on Improved YOLOv5
Tianyu Gao, Mairidan Wushouer, Gulanbaier Tuerhong · 2022
An improved small object detection method based on YOLOv5 algorithm is presented to solve the problems of dense and uneven small target samples, less extractable feature information and susceptible to background interference in aerial images taken by unmanned aerial vehicles. Firstly, CBAM attention mechanism is introduced into the network. Secondly, a new small object detection layer is added to realize four detection structures to recognize different size objects, to increase the ability to detect small and weak objects. Finally, in the post-processing section, EIoU_Loss is used to instead GIoU_Loss as a loss function of bounding box regression in order to increase the speed of bounding box regression and increase positioning precision. Experiments on the public dataset VisDrone, demonstrate that this method's average accuracy is 36.7%, which is 7.6% more effective than the standard procedure. The results show that the method proposed in this paper has good performance for UAV small object detection tasks.