To address the challenges of morphological diversity, multi-scale feature coupling, and complex noise interference in structural plane recognition from borehole images during deep-buried tunnel engineering, the authors propose an enhanced U-Net-based deep learning model for automated semantic segmentation and quantitative geometric characterization of structural planes. Targeting the limitations of fixed receptive fields in conventional convolutional networks, a dynamic snake convolution module is integrated into the encoder path to enhance geometric feature extraction capability for irregular structural planes through adaptive deformable sampling. The decoder path incorporates an efficient channel attention mechanism to improve sensitivity to critical channel features. Utilizing a selfconstructed borehole image dataset (770 images) with a two-phase training strategy, experimental results demonstrate that the model achieves a mean intersection over union of 68.44%, pixel accuracy of 93.94%, and segmentation precision of 86.12% for structural planes on the validation set. Furthermore, a sinusoidal curve fitting algorithm is implemented to automatically extract orientation, dip angle, and width parameters from segmentation results, yielding an average error below 2% compared to manual measurements. By synergizing deep learning with geometric modeling, the proposed approach remarkably enhances the accuracy and efficiency of structural plane recognition under complex geological conditions, providing reliable technical support for quantitative rock mass stability analysis, and further promoting the paradigm shift in geological exploration from empirical judgment to intelligent analysis.