Abstract:To improve the recognition rate of similar abrasive particles in ferrous abrasive particles,reduce the missed detection rate of small particles,and ensure the realtime detection speed,two improved models,large detection region of yolo layers(yolov 3_mod) and fully detection region of yolo layers(yolov 3_5 l),were proposed based on the YOLO algorithm.The two improved models increase the recognition rate of similar abrasive particles and reduce the missed detection rate of small particles by adding the Spatial Pyramid Pooling Module (SPP Module) and expanding the scale of the yolo layer to improve the backbone network structure,and reduce the amount of model calculation and increase the speed of model detection by merging the convolutional layer with batch normalization (BN) layer.The experimental results show that compared with the original model,the recognition rate of yolov 3_mod for similar abrasive particles is increased by 8% and the mean average precision is increased by 5%,and the recognition rate and mean average precision of yolov 3_5 l are increased by 14% and 10%,respectively.Compared with the original model,the inference speed of the two improved models is increased by 8%,and the models have better positioning accuracy,which basically realize the recognition of multitarget wear particles in complex background.The yolov 3_mod model has a faster detection speed,and the yolov 3_5 l model has a higher detection accuracy,which can be selected based on actual working conditions.