Segmentation network of concrete cracks with multi-frequency OctaveRes dual encoder and cross-attention mechanism optimized by average weight

Abstract

Concrete cracks are one of the most harmful flaws on the road, threatening traffic safety. In this paper, an effective crack segmentation network MOACA-CrackNet that strives to boost both the model generalization rate and segmentation accuracy of crack segmentation is proposed to segment various types of cracks rapidly and accurately in a variety of acquisition conditions. First, a multi-frequency OctaveRes dual encoder is designed to reduce spatial redundancy by sharing information from neighboring locations. Then, an average weight cross-attention mechanism is designed to suppress redundant background information and improve information exchange between frequencies. Finally, depthwise separable convolution is used to reduce the number of parameters. A dataset with a total of 2062 crack images is constructed in this research, MOACA-CrackNet is trained and tested on this dataset. The experimental results show that MOACA-CrackNet has a good segmentation performance for tiny cracks, the F1-score and mIoU reached 89.2% and 91.32%, respectively.