The Beishan Underground Research Laboratory (URL) is an underground research laboratory that studies the scientific topics of the high-level radioactive waste (HLW) repository in China. The geological formation of the Beishan URL is crystalline granite. The inherent joints in the granite formation possibly degrade the integrity and stability of the HLW repository; thus, it is significant to investigate the detailed geological formation joint information of the URL. This paper presents the rock outcrop joint identification research of the Beishan URL based on deep learning. The hybrid domain neural network attention mechanism was added to the encoder network module of the DeeplabV3+ +. An enhanced rock joint identification neural network was built, and a rock outcrop joint training dataset was created. The rock joint identification neural network was trained using the TensorFlow deep learning framework, which is used to obtain the rough identification results of the rock joints from sub-images. The joint skeleton algorithm based on the intersection and union ratio was used to remove the fake branches of the rock joint identification results based on deep learning. A long-joint connection algorithm is proposed to prevent long joints running through sub-images from being wrongly segmented. This algorithm effectively connects long joints between sub-images and avoids the problem of long joints in large images being truncated. The coordinates of key points of joints were reconstructed through the pixel-matching method from the sub-images, and the digitalized overall joint network of typical outcrops was constructed.