Abstract: Coal-rock identification is the core technology to realize an intelligent coal mining, which has become a technical problem in the field of coal mining.In view of the low accuracy of coal-rock identification in the fully mechanized face, a coal-rock image recognition method based on improved U-net network model is proposed.This method reduces the parameters of the network model by using deep separation convolution instead of traditional convolution, and improves the efficiency of semantic segmentation.The Res2 net module is also used to improve the ability of encoder to extract features.At the same time, a conditional random field is added to post-process the segmented image to improve the accuracy of the network model in the segmentation area.In order to obtain more abundant coal-rock distribution images, coal-rock samples with different characteristics are developed, and the coal-rock cutting experimental platform of shearer is built.The coal-rock images are obtained through the coal-rock cutting experiments, and some operations such as segmentation, scale, rotation, shear and adding noise are simultaneously performed to produce a rich data set for coal-rock image semantic segmentation, containing 8 000 images.The adaptive learning algorithm is used to train the model, and the change rules of accuracy and loss function in the process of model training are given.The pixel accuracy and intersection over union are selected to evaluate the semantic segmentation results.The results show that the pixel accuracy and intersection over union of improved U-net network model are 95.81% and 91.13%,respectively.The model only occupies 35 M of memory, and the test time is 36.45 ms/sheet which is superior to other network models in the aspect of coal-rock segmentation.In the underground field experiment, the improved U-net network model is trained and tested by constructing the semantic segmentation data set of coal and rock images collected from the fully mechanized coal mining face.Finally, the coal-rock identification of fully mechanized mining face is realized, which verifies the feasibility and practicability of this proposed method.