Abstract: Traditional methods for panoramic unfolded image stitching of rock mass borehole walls suffer from insufficient robustness in establishing feature correspondences between adjacent images, as well as poor quality, limited quantity of extracted image feature points and otherc problems. Meanwhile, supervised learning methods cannot obtain sufficiently precise labeled matching point pairs. To address these issues, an unsupervised learning-based panoramic unfolded image stitching method for rock mass borehole wall is proposed. Multi-scale feature extraction was performed on two adjacent panoramic unfolded images of the rock mass borehole wall to be stitched, using a ResNet network improved with grouped convolutions. A matching degree cross-correlation calculation module was introduced to identify and align features within the feature maps, thereby determining the spatial relationships between corresponding feature maps. A global and local deformation offset calculation network module precisely aligned spatial features of the images. Furthermore, homography deformation and image grid deformation modules effectively eliminated feature distortions between adjacent images, achieving overall alignment and fine local adjustments, enabling accurate registration of local features and deformations. Experimental results showed that this method effectively overcame issues such as image feature displacement, content misalignment, loss of detailed features, and stitching failure. The stitching seams exhibited almost no visible artifacts, improving the overall quality and visual effect of the stitched images. The method outperforms other mainstream image stitching approaches in terms of Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR) in overlapping regions, and Structural Similarity (SSIM) index, significantly enhancing the stitching accuracy of panoramic unfolded images of rock mass borehole walls.