Two-step deep learning approach for pavement crack damage detection and segmentation

ABSTRACT

Crack is a common disease of pavement, which will lead to more serious problems if it is not found and maintained in time. This means that it is very important to accurately extract and measure the damage information of pavement cracks. Compared with the traditional methods, the automatic detection and segmentation of pavement cracks using visual elements are more effective which has become a focused area. Although extensive researches has used deep learning methods in pavement crack detection, these methods only involve the single task of detection or segmentation, and few research optimises and combines them. In addition, the accuracy and inference speed of pavement crack detection and segmentation algorithm is also worthy of further research. To solve these limitations, this research proposes a new method of two-stage pavement crack detection and segmentation based on deep learning. The proposed method combines pavement crack detection and segmentation. In the first stage, the optimised YOLOv4 is used as the pavement crack damage detection algorithm to detect pavement cracks under various complex backgrounds. In the second stage, the cracks detected in the first stage are segmented, the detection accuracy is specific to the damage pixels. To further optimise the performance of the detection and segmentation algorithm, a new deeplabv3+ pavement crack segmentation method based on the Ghost module and CBAM attention mechanism is proposed. Compared with the original network, the proposed two-stage pavement damage detection and segmentation method improve the detection accuracy by 2.23$\text{\%}$ and 7.47$\text{\%}$, respectively. The network inference speed is improved by 35.3$\text{\%}$ and 50.3$\text{\%}$, respectively. Compared with the existing single-stage pavement damage detection or segmentation methods, the proposed method has the advantages of fast inference speed and high detection accuracy.

KEYWORDS:

• Pavement damage detection
• deep learning
• crack segmentation
• YOLOv4
• deeplabv3+

Disclosure statement

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Additional information

Funding

The authors are grateful for the National Natural Science Foundation of China [grant numbers 62073196, 62076150], the Taishan Scholar Project of Shandong Province [grant number TSQN201812092], Natural Science Foundation of Shandong Province [grant number ZR2020MF109] and the Key Research and Development Program of Shandong Province [grant numbers 2021CXGC011205, 2021TSGC1053].