Pixelwise asphalt concrete pavement crack detection via deep learning‐based semantic segmentation method. Issue 8 (5th April 2022)

• Record Type: Journal Article
• Authors: Huyan, Ju; Ma, Tao; Li, Wei; Yang, Handuo; Xu, Zhengchao
• Is Part Of: Structural control and health monitoring. Volume 29:Issue 8(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-04-05
• Subjects: deep learning -- pavement cracks -- pixel‐level crack detection -- semantic segmentation -- U‐net; Structural engineering -- Periodicals; Structural control (Engineering) -- Periodicals; Automatic data collection systems -- Periodicals; Detectors -- Periodicals; 624.17
• DOI: 10.1002/stc.2974
• Languages: English
• ISSNs: 1545-2255

Abstract

Summary: Explicit gaps exist between the advanced deep learning technologies and the less satisfied pixel‐level crack detection algorithms. Therefore, this research sought to bridge this gap via outlining the deep neural network model for pixelwise pavement crack detection. Two state‐of‐the‐art deep neural network models are constructed for the semantic segmentation of crack images. The first architecture, VGGCrackU‐net, is composed of 10 3 × 3 convolutional layers, 4 max‐pooling layers, 4 up‐sampling layers, and 4 concatenate operations. Another architecture, ResCrackU‐net, is composed of 7‐level residual units with a total of 22 convolutional layers. Asphalt concrete pavement crack images are collected by smartphones, action cameras, and automatic pavement monitoring systems from diverse functional classes of AC pavements. The crack images are manually labeled and double‐checked by trained operators for quality insurance. After that, 500 crack images are randomly divided into training, validating, and test datasets according to the ratio of 3:1:1. Both architectures are trained on GPU facilitated Keras platform with Python version of 3.5, which demonstrated fast convergence. Results prove that the proposed models exhibit significant advantages for pixelwise crack detection when compared with the performance of widely used FCN net and PSPnet. Meanwhile, ResCrackU‐net slightly outperforms VGGCrackU‐net, which, however, can provide acceptable results as well. Though significant false negative and false positive errors are observed in both network models, the contributions are noticeable, which can provide innovative guidance for future work in figuring out solutions to the problems.