SnakeConv and SFC boosting precise segmentation on the crack of tunnel lining surface: based on DeepLabV3+ with improved Swin transformer V2

Abstract Tunnel cracks pose a significant threat to structural integrity, potentially leading to localized collapse of the infrastructure. Traditional manual crack detection methods are prohibitively expensive, highlighting the need for an efficient and accurate automatic crack segmentation model. To address this challenge, we propose a novel crack segmentation model for subway tunnel lining surface based on the DeepLabV3+ architecture. In this model, we design an improved Swin transformer V2 Base (SwinV2*) as the backbone to enhance crack segmentation performance. Considering the tubular morphology of tunnel cracks, we introduce a snake convolution module to better capture their unique features. To prevent performance degradation when fusing shallow and deep features, we incorporate a spatial feature calibration module that facilitates feature alignment and grouping along the channel dimension. We assess our model’s effectiveness using thousands of crack images captured by the image acquisition system designed for subway tunnel surfaces. Experimental results show that our model achieves strong performance metrics: 68.96% IoU, 84.33% mIoU, 87.57% PA. Compared to the original DeepLabV3+, our approach demonstrates superior performance, with a 2.89% improvement in IoU, a 1.45% increase in mIoU and notably, a significant 10.39% improvement in PA.