MCL-CrackNet: A Concrete Crack Segmentation Network Using Multilevel Contrastive Learning
Pengfei Shi, Shen Shao, Xinnan Fan, Zhongkai Zhou, Yuanxue Xin
Abstract
Automatic concrete crack segmentation by computer vision is a challenging task due to the complex environments with uneven illumination, low resolution and excessive noise. Meanwhile, existing methods suffer from low detection accuracy and weak generalization ability. To solve these problems, we propose a novel crack detection network using multi-level contrastive learning, called MCL-CrackNet(Multi-level Constrastive Learning - CrackNet). First, we leverage a Coordinate Convolution Block for extracting the features with geometric information by adding two channels with the coordinate information of the features. Thus, the extracted features with rich geometric information can have a better spatial perception. Then, we propose a Position Attention Gate module that fuses downsampling and upsampling features of the same size, and thus makes the fused features more focused on the position information of the cracks. Finally, we obtain more useful information of local and global features with multi-level contrastive learning instead of single-level contrastive learning. Compared with other previous methods, MCL-CrackNet achieves accurate detection with better generalization ability in both underwater dam cracks and pavement crack scenarios.