Abstract: Current methods of joint information recognition are only applicable to local rock images. To address this limitation, we employed the panoramic developed imaging technique to extract image features, reconstruct point-cloud model, and correct and stitch the collected local rock images, thereby obtaining high-resolution panoramic image of the tunnel’s surrounding rock mass. Through image pre-processing and recognition of small-size feature images by SmAt-Unet neural network, followed by fusion of the recognition results, we roughly recognized the joint occurrences in the panoramic image region. Subsequently, we extract the refined joint information via skeletonization, skeleton line separation, burr removal, and skeleton line connection using the Zhang-Suen algorithm and the 8-neighborhood connected domain analysis method. Ultimately, through quantified analysis of volumetric joint number and joint occurrence information, we developed the method to identify rock joint information based on panoramic developed images. Application results demonstrate an average fitting error of 0.90 of the spatial equation of jointed plane, indicating successful joint information identification. Moreover, the panoramic developed imaging technique boasts advantages such as rapidity, simplicity, and flexibility, with minimal impact on site construction.

Key words: tunnel, surrounding rock, panoramic developed image, SmaAt-Unet neural network, joint recognition