APLCNet: Automatic Pixel-Level Crack Detection Network Based on Instance Segmentation

• Main Authors: Yuefei Zhang, Bin Chen, Jinfei Wang, Jianming Li, Xiaofei Sun
• Format: Article
• Language: English
• Published: IEEE 2020-01-01
• Series: IEEE Access
• Subjects: Pavement crack detection; deep learning; object detection; semantic segmentation; instance segmentation
• Online Access: https://ieeexplore.ieee.org/document/9239304/

The accurate and automatic detection of pavement cracks is essential for pavement maintenance. However, automatic crack detection remains a challenging problem due to the inconspicuous visual features of cracks in complex pavement backgrounds, the complicated shapes and structures of cracks, and the influences of weather changes and noise. In recent years, with the development of artificial intelligence technology, crack detection methods based on classification and semantic segmentation have laid a good foundation for the automation of pavement crack detection. However, there remain shortcomings in the comprehensive acquisition of pavement crack attribute information and detection accuracy. To solve these problems, this paper proposes an instance segmentation network for pavement crack detection. The network can simultaneously obtain the crack category, position, and mask, and can realize end-to-end pixel-level crack detection. A semantic segmentation branch is first added to Mask R-CNN. This branch can extract the bottom-level detail information of the cracks and ultimately improves the accuracy of crack mask prediction. An adaptive feature fusion module is then designed. During feature fusion, this module highlights the attribute information and location information of cracks according to the channel attention mechanism and the spatial attention mechanism. Finally, these two modules are integrated to form an automatic pixel-level crack detection network, namely APLCNet. Without any embellishment, APLCNet achieves a precision of 92.21%, a recall of 94.89%, and an F1-score of 93.53% on the challenging public CFD dataset, thereby outperforming CrackForest and MFCD for pixel-wise crack detection. Moreover, APLCNet achieves a 16.5% mask AP on the self-captured GDPH dataset, thereby surpassing Mask R-CNN and PANet.