An improved U-Net based method for predicting cable tunnel cracks

To rapidly and accurately predict cable tunnel crack, a method based on improved U-Net model was proposed in this research. The new semantic segmentation network called IRHC U-Net is developed based on U-Net, which effectively integrates Inception-residual convolution (IRC) module, hybrid attention (HA) module, hybrid dilated convolution (HDC) module, and cross-level gathering decoder (CLGD) module. The IRC module can increase the network depth and width, and solve the problem of the gradient decreasing; The HA module highlights the cracks of edge features; The HDC module can obtain the more detail information in the en-coder; The CLGD module narrows the semantic gap between encoder and decoder. The prediction results of cable tunnel cracks showed that the improved U-Net model outperformed the other models. The mean intersection over union (MIOU), Accuracy, and Precision of the improved U-Net model on the cable tunnel cracks dataset were 84.79%, 99.82%, and 91.35%, respectively. The results indicate that the improved U-Net model can accurately predict cable tunnel crack. This study provides technical support for the rapid and accurate segmentation of cable tunnel cracks. • The IRHC U-Net achieved MIOU, Accuracy, and Precision of 84.79%, 99.82%, and 91.35%, respectively, outperforming other models on the self-built cable tunnel crack dataset. • The integration of IRC, HA, HDC, CLGD, and other modules enhances the model’s adaptive learning and feature detail extraction capabilities, allowing for accurate and efficient segmentation of crack features in the cable tunnel environment. • The IRHC U-Net demonstrates strong generalization and robustness across various fracture datasets, making it effective for segmenting fracture-related data.