Small Targets Under Drone Perspective Based on Improved YOLOv5s Detection Algorithm

Research aimed at the detection of ground targets using UAVs for shooting, the phenomenon of missed detections and false detections still exists, and the algorithm cannot simultaneously consider detection accuracy and parameter size. Therefore, this paper uses the improved YOLOv5s algorithm to further study the above problems. Firstly, a Simple Patches of Convolution-Mixer (CSPCM) is added to both the Head and Backbone network structures of the original model, which uses deep convolution to help better fuse the information of the two dimensions and achieve higher accuracy; In addition, a small target detection layer is set up in the neck to expand the perception field of the model, so that the algorithm model can be handy when dealing with small target detection tasks. Secondly, the lightweight convolutional network module GhostNetV2 is used as the backbone of the algorithm, reducing feature redundancy in channels. Finally, YOLOv5-CG was tested on VisDrone2019. The new model reduced the number of parameters by 1468493 and the model size by 2.4M, mAP@0.5 Improved by 4.1%, mAP@0.5 0.95 increased by 1.9%; The experimental results indicate that Yolov5-CG can achieve parameter simplification while ensuring improved accuracy, and can be better used for small object detection.