LinkCrack_IDL: a pixel-level tunnel crack segmentation algorithm based on deep learning

Abstract As a key facility in modern transportation networks, detecting fine cracks in tunnels is crucial in disaster prevention. However, segmenting fine cracks is an open issue in complex contexts. To address this issue, a deep learning-based model LinkCrack_IDL is proposed in this paper, which designs information retention downsampling module (IRDM), depthwise separable feature enhancement module (DSFEM), and learnable upsampling module (LUM). First, for fine crack features in complex backgrounds, the information loss caused by traditional downsampling will directly affect the model’s segmentation accuracy. IRDM dynamically adopts different downsampling strategies through the number of channels to retain more features. Second, traditional convolution is computationally intensive, and the high-frequency features tend to cover the information of fine crack edges, which makes it difficult for the model to extract the crack features effectively. DSFEM adopts multi-scale feature extraction combined with depthwise separable convolution to capture the fine edges of the cracks and the texture features in the complex background. Then, conventional upsampling methods do not have learning capability in crack detection models, which affects resolution recovery and crack segmentation accuracy. LUM enhances the detail recovery capability through the primary and secondary path structure while optimizing the feature representation so that the model can segment the crack region more accurately. Finally, comparative experiments show that the proposed algorithm LinkCrack_IDL outperforms other methods in terms of performance. Especially on real tunnel scanning images, the performance indicators F1, IoU, and Precision of crack detection have reached 77.41%, 63.68%, and 81.26%, respectively, which are at least 6.93%, 8.59%, and 7.83% higher than other methods. In conclusion, the proposed algorithm LinkCrack_IDL provides an advanced approach for tunnel crack segmentation, which can facilitate early detection of potential structural failures, ensuring timely repairs and reducing maintenance costs in critical transportation networks.