High-precision segmentation and quantification of tunnel lining crack using an improved DeepLabV3+

Abstract

Current semantic segmentation models have limitations in addressing tunnel lining crack, such as high complexity, misidentification, or inability to detect tiny cracks in specific practical scenarios, which is crucial for precise assessment of tunnel lining health. We developed a novel approach called EDeepLab, aiming to achieve a higher precision detection and segmentation of lining surface crack. EDeepLab improves upon the original DeepLabV3+ framework by replacing its backbone network with an optimized lightweight EfficientNetV2. The amount of EfficientNetV2 block computation is reduced and a self-designed shallow feature fusion module is used to merge the layers to enhance parameter utilization efficiency. Furthermore, the normalization-based attention module and convolutional block attention module attention mechanisms are integrated to classify and process both high and low dimensional information features. This allows for comprehensive utilization of global semantic information and channel information, thereby enhancing the model’s feature extraction capability. Results in constructed metro-tunnel crack dataset demonstrate that the number of parameters is reduced from 144.45 M in the DeepLabV3+ to 99.80 M in the EDeepLab. EDeepLab achieves a mean intersection over union of 84.77%, mean pixel accuracy of 94.96%, and frames per second of 18.52 f/s. The proposed EDeepLab outperforms other models including U-Net, ResNet and fully convolutional networks in the quantitative analysis of tiny cracks and noise interference.