Pile Foundation Borehole Crack Detection Method Based on Deep Residual Network Guided by Dice Loss

As a critical load-bearing structure in civil engineering, the quality inspection of pile foundations is of paramount importance. Traditional core drilling methods encounter challenges of missed detection and misjudgment in identifying crack defects within pile shafts. Borehole television imaging technology provides authentic and intuitive image data of internal concrete structures, effectively addressing these limitations. However, current applications of borehole television imaging still face efficiency and accuracy challenges in pile quality assessment. To enhance the performance of borehole television imaging in crack detection, this study proposes an intelligent crack recognition method based on deep residual semantic segmentation for borehole television images of pile foundations. The method employs a deep residual convolutional neural network model integrated with advanced image processing and analysis techniques, achieving automated identification of crack features in borehole television images. Through targeted improvements to the loss function, the approach significantly reduces omission rates in detecting fine cracks. Experimental results demonstrate that the proposed method achieves remarkable accuracy (82.2%) and robustness in identifying pile shaft crack defects. The system effectively recognizes and annotates cracks in borehole television image data while substantially improving detection efficiency and reducing inspection costs. This intelligent crack recognition methodology based on deep residual semantic segmentation demonstrates superior performance in pile foundation crack detection, effectively enhancing inspection accuracy and operational efficiency while minimizing diagnostic errors. The approach establishes a novel intelligent solution for pile quality assessment with extensive application prospects.