Boulder yield is an important index to evaluate the blasting quality in the blasting process of an open pit mine. Since a high boulder yield will not only greatly reduce the mining efficiency, but also increase the cost of secondary rock breaking, so fragments size statistics is an important work in open pit mining. Aiming at the problem that the statistics of fragment size is complex and not accurate enough, a statistical model of boulder yield was built by deep learning based on the takes the image data of blasting piles collected in the Unugetushan copper and molybdenum mine. Firstly, the annotated data set was initially segmented into an initial effect diagram of the mine rock contour based on the U-net image segmentation model. And then, the annotated data for training was optimized and the Resu-net model was improved on the basis of the residual learning module, which resulted in the final segmentation effect map of mine rock contour. Finally, the fragment size information of the blasting pile was obtained through the minimum external rectangle method combined with OpenCV image processing technology. The results show that the segmentation accuracy of U-net+Resu-net fragment size optimization model proposed in this study is 97.84% with an accurate image data segmentation. The statistics of fragment size in an inclined blasting pile is realized by OpenCV technology combined with the camera monocular imaging principle. In addition, the developed interactive interface is simple to operate and can quickly calculate the boulder yield.