Research on coal and gangue detection algorithm based on improved YOLOv5s model

Abstract

In order to solve the problems of slow detection speed and low detection precision of the existing deep learningbased coal and gangue target detection methods, an improved YOLOv5s model is proposed and applied to coal and gangue target detection. The YOLOv5s model is improved by embedding selfcalibrated convolutions (SCConv) in the Backbone area of YOLOv5s model as the characteristic extraction network, which can better fuse multiscale characteristic information. Because the size of coal and gangue is too small compared with the whole image, the Neck area of YOLOv5s model is appropriately simplified, and the 19×19 characteristic map branches suitable for detecting larger size objects are deleted, thus reducing model complexity and improving the realtime detection performance. The anchor box obtained by clustering with Kmeans algorithm is linearly scaled to improve the model detection precision. The experiment of coal and gangue target detection based on improved YOLOv5s model shows that compared with YOLOv5s model, the improved YOLOv5s model can detect the corresponding coal and gangue accurately. The size of improved YOLOv5s model is reduced by 1.57 MB, the frame rate is increased by 2.1 frames/s, and the average precision is improved by 1.7%, indicating that the improved YOLOv5s model has improved both detection precision and detection speed.

Keywords

• coal and gangue separation； coal and gangue identification； coal and gangue target detection； yolov5s; selfcalibrated convolutions