A spatial-channel hierarchical deep learning network for pixel-level automated crack detection

• Main Authors: Pan, Yue, Zhang, Gaowei, Zhang, Limao
• Other Authors: School of Civil and Environmental Engineering
• Format: Article
• Language: English
• Published: 2022
• Subjects: Engineering::Civil engineering; Computer Vision; Automated Crack Detection
• Online Access: https://hdl.handle.net/10356/161075
• Institution: Nanyang Technological University

This research develops a novel computer vision approach named a spatial-channel hierarchical network (SCHNet), which is feasible to support the automated and reliable concrete crack segmentation at the pixel level. Specifically, SCHNet with a base net Visual Geometry Group 19 (VGG19) contains a self-attention mechanism, which is realized by three parallel modules, including the feature pyramid attention module, the spatial attention module, and the channel attention module. It can not only consider the semantic interdependencies in spatial and channel dimensions, but also adaptively integrate local features into their global dependencies. The segmentation performance is evaluated by a metric named Mean Intersection over Union (IoU) in a public dataset containing 11,000 cracked and non-cracked images with a unified resolution at 256 × 256 pixels (px). The experimental results confirm the effectiveness of the three attention modules, since they can individually increase Mean IoU by 1.62% (74.16%–72.54%), 5.15% (79.31%–74.16%), and 5.76% (79.92%–74.16%), respectively. With the help of new strategies like the data augmentation and multi-grid method, SCHNet can boost Mean IoU to 85.31%. In a comparison of the state-of-the-art models (i.e. U-net, DeepLab-v2, PSPNet, Ding, Dilated FCN) on the test dataset, SCHNet can outperform others with an improvement of at least 7.51% in Mean IoU. Moreover, SCHNet is robust to noises with a better generalization ability under various conditions, including shadows, roughness surfaces, and holes. Overall, this research contributes to developing SCHNet to integrate spatial and channel information in feature extraction, resulting in a more accurate and efficient crack detection process.