Discontinuities are vital components of rock mass, significantly affecting its strength, deformation, and seepage characteristics. They provide essential parameters for rock mass classification and engineering design. Borehole television technology is a widely used method for capturing these discontinuities within the rock mass, offering high-resolution in situ images. However, in complex strata, the discontinuities appear in various morphologies with significant width differences. Coupled with the rapid texture changes on rock wall faces, these discontinuities create a highly uneven contrast, making it challenging for traditional algorithms to recognize them accurately. To address this challenge, this study introduces an improved deep learning network model specifically designed for borehole images of complex strata.
The proposed model, based on the U-Net architecture, incorporates a deeper encoding-decoding network structure. This structure effectively handles semantic information related to discontinuity breaks caused by uneven contrast. The model integrates channel attention mechanisms and residual modules, enhancing feature extraction capabilities at different levels in the encoding stage. In addition, the channel attention mechanism fuses multichannel discontinuity information from both encoding and decoding layers. A multiscale spatial attention module introduced in the lower-level skip connection improves the ability to process complex morphological discontinuities and enriches the semantic features of discontinuities in the coding layer. In this study, the borehole image data are augmented in various ways, such as using perspective deformation similar to the stratum deformation under geological action. This study also employs joint training with focal loss and Dice loss to handle imbalanced image data. The generalization ability of the network model is thoroughly validated through ablation studies and comparative experiments using the same borehole image as the training set and neighboring borehole images as the test set. For comprehensive quantitative evaluation, this study uses several metrics, including precision, recall, F1-Score, and F2-Score.
Our experimental evaluation, conducted on a self-made borehole image dataset, indicated that compared to several common image segmentation network models, our proposed model significantly improved the recognition capability of rock discontinuities in borehole images from complex strata while ensuring faster computational efficiency. The precision and recall on the test set for the proposed model reached 78.23% and 77.85%, respectively. This marked an improvement in segmentation performance by 7.96% and 14.99%, respectively, compared with the basic U-Net model. Both the F1-Score and F2-Score were close to 78%. Although the model size was 18.13 MB and had approximately twice the parameters of the base U-Net, the deeper network hierarchy reduced the number of channels of shallow high-resolution feature maps, resulting in a reduction in computational load. The model achieved an FPS of 85, which was slightly higher than that of the basic U-Net model.
This study meticulously improves upon the basic U-Net model by strategically incorporating the attention mechanism, residual connections, and multiscale convolutions. The improved model exhibits high accuracy and robustness. It effectively confronts the challenges associated with balancing detailed features and high-level semantics owing to significant width differences in discontinuities within complex strata. Furthermore, it addresses issues related to incomplete extraction of discontinuities caused by uneven contrast between discontinuities and rock wall surfaces. As such, this improved model provides strong technical support for the automatic identification of rock discontinuities in on-site borehole investigations.
京公网安备11010802044758号