Aiming at the characteristics of shale bedding structures and fine fracture developments, an automatic multi-scale fracture recognition method for shale outcrop images based on a UU-Net (multi-path U-shaped network) deep neural network was presented. The process begins with the manual labeling and binarization of fracture image datasets. The datasets were then divided into training and testing subsets. Through iterative parameter optimization, the model was trained and validated to achieve the desired accuracy, resulting in a multi-scale fracture recognition model for outcrop images. The model was subsequently applied to shale outcrop images and rock slice images from the Sichuan Basin to extract fractures at different scales. Quantitative statistics on fracture length and dip angle information were provided. The application results demonstrate as follows. Compared to manual labeling, the efficiency of crack recognition using deep learning is significantly improved. The accuracy of large-scale fracture identification is 98%, while the accuracy for medium- and small-scale fractures is approximately 96%, and for micro-fractures, it is about 94%. The trained model can be applied to multi-scale fracture identification of shale outcrops or rock slice images in other regions, indicating strong generalization capabilities.