This paper is the result of geological survey engineering.

Objective The precise identification and extraction of rock fractures are of great significance for geothermal resource development, rock mass seepage analysis, and related underground engineering, and they form the basis for studying the mechanical properties and seepage characteristics of rock masses. However, due to the complex imaging environment, diverse fracture morphologies, and background texture interference, rock fracture images often have problems such as thin fracture areas, low contrast, and significant noise, making it difficult for traditional image segmentation methods to achieve ideal results in feature extraction and multi-scale information fusion. Therefore, this paper aims to improve the accuracy and robustness of rock fracture extraction and proposes a rock image fracture extraction algorithm based on the improved DeepCrack - ResCrack.

Methods This paper makes improvements from two aspects: network structure and feature fusion mechanism. Firstly, a new backbone network structure is designed, introducing deeper residual connections to enhance gradient propagation and feature reuse capabilities, replacing the original VGG backbone, thereby more fully extracting multi-scale fracture features. Secondly, a strip pooling module (SPM) is introduced, which performs long-range context modeling along the horizontal and vertical directions, enhancing the model's sensitivity to fracture structures and effectively suppressing the interference of background noise.

Results The proposed ResCrack algorithm demonstrates excellent performance on the self-built rock fracture image dataset. Compared with the original DeepCrack model, this method improves the mean intersection over union (mIoU), mean pixel accuracy (mPA), and F1 score by 1.88%, 4.49%, and 3.02% respectively, and can extract continuous fracture structures more completely. At the same time, it shows stronger anti-interference ability in complex background areas.

Conclusions The ResCrack algorithm proposed in this study can effectively enhance the recognition accuracy and robustness of rock fractures. It not only performs well on experimental datasets but also demonstrates excellent application potential in real geothermal scenarios. It can provide an effective basis for infering the fracture structure of underground reservoirs and has positive significance for promoting the in-depth application of computer vision in the field of geoscience.