Asphalt Pothole Detection via Grayscale-Texture Fusion and Fast R-CNN with Morph Postprocessing

Abstract

Aiming at the problems of low recognition accuracy and insufficient feature extraction in existing vision-based pothole detection methods, this paper proposes a multi-stage detection method that integrates grayscale and texture features. The method flow consists of four stages: firstly, the pavement image is binarized to achieve pothole qualitative identification (92.2% accuracy) and coarse extraction based on shape features and local standard deviation; secondly, texture features of the candidate region are extracted through the grayscale covariance matrix and principal component analysis (PCA) is used to eliminate the feature redundancy; then, the texture region that conforms to the pothole lesion characteristics is aggregated with the results of the pothole lesion detection using fuzzy C-means clustering algorithm. characteristics of the texture regions are aggregated and spatially superimposed with the coarse extraction results; finally, the boundary is optimized by morphological post-processing to obtain accurate pothole segmentation results. On the dataset containing disturbing scenes such as cracks, gravels, water accumulation, etc., the method achieves a recall rate of 90.0% and a precision rate of 87.1%, in which 70.4% of the samples have an intersection and union ratio (IoU) of more than 80%, and 85.2% of the samples have an IoU of more than 70%. Experiments show that the detection performance of this method in complex pavement environments is significantly better than that of traditional single-feature detection models.