This study focuses on enhancing crack identification in rock surfaces using deep learning techniques. The research proposes a novel convolutional architecture to achieve pixel-level classification for distinguishing cracks from backgrounds in high-resolution images. Key innovations include the removal of pooling layers for precise segmentation, the incorporation of deconvolutional layers to maintain spatial resolution, and addressing challenges like class imbalance with data augmentation and adjusted loss functions. A key challenge is the scarcity, low quality, and class imbalance of training data—only 4 8-bit grayscale images (0–255 grey levels) with high noise (up to dozens of grey levels) are available. Tested on these manually annotated datasets, the proposed model achieves reliable accuracy and scenario-specific robustness in ESEM mudstone microcrack segmentation, outperforming conventional methods and enabling real-world applications in geology and structural monitoring. • A novel deep learning architecture for vision based rock crack detection. • Address image class imbalance with data augmentation. • Outperform conventional image segmentation model.
Lu et al. (Sun,) studied this question.