Abstract: Coal rock fractures are closely related to gas migration and affect the stability of coal rock. Studying the complex fracture system in coal rock is of great significance for roadway support and gas extraction. At present, the recognition methods for coal rock fracture images fail to comprehensively consider the features of the number, position, morphology, and category of fracture in coal rock images, making it difficult to obtain effective information. Taking the coal rock images of excavation face in the No.8 Coal Mine of Hebi Coal and Electricity Co., Ltd. as the research object, a pixel level intelligent recognition method based on U-Net network for coal rock fractures and categories is proposed. The histogram equalization, Gaussian bilateral filtering, and Laplace operator are used to preprocess coal rock images to improve image quality and extract fracture feature information more effectively. The features of coal rock fractures are recorded by observing and divided into 7 categories, the selected coal rock fracture images are amplified, and the images are annotated at the pixel level using Labelme software to establish a coal rock fracture dataset. The U-Net network is used to construct a coal rock fracture recognition model. After debugging, the network batch size and learning rate parameters are determined. The experiment shows that when the number of iterations reaches 300 or more, the average recognition accuracy of the model is 87%, the average recall rate is 92%, the average intersection to parallel ratio is greater than 85%, and the average pixel accuracy of the category is greater than 80%. The coal rock fracture recognition model is validated by collecting underground coal rock mining fractures and laboratory tensile exogenous fractures. The results show that the model can effectively extract target feature information and distinguish it from background feature information, and can accurately locate and recognize a single fracture.