Abstract: The identification and measurement of rock joint cracks is an important basis for the characteristics of rock structural planes and the evaluation of rock quality. The traditional identification of rock joint cracks in boreholes is mainly manual and inefficient. In this paper, the deep learning convolution neural network Mask R-CNN algorithm model is used to construct the data set by labeling and standardizing the features of the image joints and cracks. The Mask R-CNN model is trained to construct the intelligent recognition and segmentation model of rock joints and cracks, so as to realize the intelligent recognition and segmentation of the observed image. The rock joint cracks are skeletonized by binarization and sine wave fitting, and the occurrence of rock joint cracks is calculated, so as to realize the statistics of rock joint cracks on the borehole wall by digital means. The technology is compared and verified by project examples. The results show : ①the non-directional rock joint fracture can be extracted by a specific algorithm model, and its corresponding occurrence can be calculated by multi-point fitting of sine curve. ②The accuracy of the measurement results can reach more than 85 %. ③With the increase of the number of different morphological samples of joint cracks, the recognition accuracy can be increased. ④The intelligent identification and measurement results of rock joint cracks can describe the rock structural plane in detail and be applied to the evaluation of rock integrity and quality grade.