A Semantic Segmentation Algorithm of Cracks in Tunnels Based on Dual-branch and Attention Fusion

In order to address the issue of low segmentation accuracy of tunnel cracks in complex environments, a Feedback Connection Module (FCM) and a Channel Attention Fusion Module (CAF) are proposed to enhance the U-Net neural network. By designing a dual-branch structure, a new feature extraction network is introduced in the encoder with feedback-type skip connections to improve the feature extraction capability of network. Additionally, a Channel Attention Fusion Module is embedded in the decoder to perform feature fusion based on the importance of feature information, thereby improving the accuracy of crack segmentation. Moreover, VGG16 is used as the backbone network for U-Net. Experimental results show that, compared to the original network, the IoU index is increased by 4% and the F 1 index is improved by 3.9%. This method also demonstrates excellent performance compared to other mainstream approaches and provides strong support for applications and further exploration in related fields.