Research on Tunnel Crack Identification Localization and Segmentation Method Based on Improved YOLOX and UNETR++

To address the challenges in identifying and segmenting fine irregular cracks in tunnels, this paper proposes a new crack identification, localization and segmentation method based on improved YOLOX and UNETR++. The improved YOLOX recognition algorithm builds upon the original YOLOX network architecture. It replaces the original CSPDarknet backbone with EfficientNet to enhance multi-scale feature extraction while preserving fine texture characteristics of tunnel cracks. By integrating a lightweight ECA module, the proposed method significantly improves sensitivity to subtle crack features, enabling high-precision identification and localization of fine irregular cracks. The UNETR++ segmentation network is adopted to realize efficient and accurate segmentation of fine irregular cracks in tunnels through global feature capture capability and a multi-scale feature fusion mechanism. The experimental results demonstrate that the proposed method achieves integrated processing of crack identification, localization and segmentation, especially for fine and irregular cracks identification and segmentation.