Tunnel Crack Detection With Linear Seam Based on Mixed Attention and Multiscale Feature Fusion

Crack detection techniques have been rapidly developed in recent years due to the rise of deep learning. However, existing methods struggle to produce accurate crack segmentation results because cracks and linear seams on the tunnel lining surface have significant similarities in terms of intensity value and texture features. At the same time, due to the scarcity of data, the existing tunnel lining surface crack detection methods still use multi-step traditional image processing methods for detection, which is inefficient. In this paper, we collect and label a dataset of 200 tunnel lining surface crack images named Tunnel200. For the first time, a deep learning-based method is used to detect cracks in the tunnel lining surface. To deal with the characteristics of crack and linear seam, which mostly present long strip or curved shapes, we propose a Mixed Attention (MA) module by efficient embedding channel and positional information. Unlike common spatial attention that aggregates information throughout space, mixed attention aggregates feature directly along with two directions, height, and width, in the spatial dimension. In this way, the long-range dependence of the crack features can be effectively captured. The proposed MA is simple to incorporate into the network. Meanwhile, we embed it in the traditional U-shape network while employing an efficient multi-scale feature fusion technique to build the Tunnel Crack Detection Network (TCDNet). TCDNet outperforms other crack detection and semantic segmentation methods on the Tunnel200 dataset. Additionally, we evaluate our method on two publicly available crack datasets, Crack500 and DeepCrack, and our method gets superior performance.