Automatic segmentation of tunnel lining defects based on multiscale attention and context information enhancement

Abstract

To address the dual challenges of complex environmental interference and multiscale targets in deep learning-based tunnel lining defect identification, a novel segmentation algorithm, named Tunnel Lining Defect Segmentation with Multiscale attention and Context information enhancement (MC-TLD), is proposed. In MC-TLD, a context-enhanced feature encoder is designed to compensate the deficiency of Convolutional Neural Network to extract global context information fully, thus reducing the misdetection and omission caused by complex environmental interference. In addition, an atrous spatial pyramid pooling module with multiscale attention is designed to improve the feature extraction capability of the network for different scales of defects. Finally, the traditional linear interpolation upsampling method in the feature decoding module is replaced by parameter learnable DUpsampling to output more accurate pixel prediction results. The experimental results show that MC-TLD has higher segmentation accuracy than other comparison models, which is suitable for the task of tunnel lining multiscale defect identification in complex environments.