Abstract: To improve the accuracy of coal rock recognition, this study collected the original images of coal rock from the excavation face in Yushujing coal mine of Shanghai Temple Mining Co. Inner Mongolia, and produced a deep learning dataset. The dataset is trained by three kinds of network models, including FCN fully convolutional neural network (FCN network), U-net Semantic Segmentation Network (U-net Network), and U-net Network improved by adding Canny Edge Detection Algorithm, and the training results were compared and analyzed. The results show that the accuracy of the three network models is 89.25%, 93.52% and 94.55%, respectively. When the number of training times reaches 100, the accuracy of the improved U-net network model increased by 1.03%. In coal rock identification, the U-net network model achieved higher accuracy than the FCN network model and showed better performance in the testing session. In the prediction session, the recognition of the edge part of the coal rock was achieved with more accurate treatment. The method can provide a reference for improvement of the accuracy of coal rock recognition.