Pixel-level Road Crack Detection and Segmentation Based on Deep Learning

Abstract

<p>This paper proposed an integrated framework for detecting and segmenting road cracks in complex backgrounds. Based on the latest real-time object detection algorithm, YOLOv5l6, a modified U-Net embedded Bottleneck and Attention mechanism modules was developed to segment crack pixels from the detected crack regions. Validation of the proposed approach was conducted based on a total of 150 images, which were taken from different backgrounds, angles, and distances. Based on the computation, the results derived from the YOLOv5l6-based crack detection had a mean average precision of 92%, and the mean intersection of the union of the modified U-Net was 87%, which is at least 11% higher than the original U-Net model. The results showed the integrated approach could be a potential basis for an automated road-condition evaluation scheme for road operation and maintenance.</p>