Abstract: Aiming at the problem that the ore fragmentation of the conveyor belt needs to be manually measured,a method for detecting ore fragmentation based on deep learning is proposed.This method adopts the residual neural network structure under the Darknet framework to form the main feature extraction network of CSPDarkNet21,considering that only large pieces of ore need to be identified and judged,simple two-way feature fusion PANet is selected as the feature extraction network and PANet is simplified from three feature layers to one feature layer,it speeds up model training and prediction,and uses CIOU to calculate the loss value to make training more stable.After the ore identification is completed,the ore fragmentation judgment structure is used to calculate the pixel area of the prediction box to obtain the true size of the ore.In the final test,the test is performed after the image is processed by grayscale and median filtering.The results show that compared with the single UNet image segmentation algorithm for judging the ore size,the IOR method is reducing the accuracy by 6 percentage points.It reduces model training time by 2.7 times and improves model operation efficiency by 9.07 times.It is a method that can quickly train and predict the degree of ore crushing.It is very suitable for use in on-site environments such as conveyor belts that require rapid identification and judgment.At the same time,it reduces the cost of equipment requirements and training costs of cloud computing.