Crack Detection in Borehole-wall Panoramic Images using Deep Learning

In geological surveys, a borehole camera is used to photograph the vertical cylindrical borehole-wall to investigate underground cracks. Currently, the identification of cracks from borehole-wall images is performed visually by skilled workers, which requires a great deal of time and effort. In this study, we use deep learning to detect sine-curve-like cracks from borehole-wall panoramic images. We designed a two-class classification model that discriminates the presence or absence of cracks using existing network architectures. Furthermore, we introduced a new data augmentation technique called "CyclicShift", which takes advantage of the unique properties of borehole-wall panoramic images. Through experiments using our own dataset, we showed that both WideResNet and ViT achieve over 98% accuracy under the limited condition of a single crack in one image. Additionally, we confirmed the effectiveness of data augmentation and fine-tuning of pre-trained models. We also demonstrated the potential of using Grad-CAM to locate the positions of cracks.