Crack Identification and Quantification of Bridge Concrete Based on YOLOX and Image Processing Techniques

Abstract

The investigation and analysis of bridge distress are critical for the assessment and maintenance of bridge safety, necessitating precise information regarding the condition of the bridge surface. In this study, a deep learning framework for automatically identifying bridge concrete cracks is proposed based on comparing the detection performance of YOLOX, SSD, and Faster R-CNN. The deep learning model YOLOX_s is initially trained and employed to identify bridge concrete cracks, and the detection results demonstrate that the bridge concrete crack identification accuracy rate of the YOLOX_s is 91.77% and much higher than that of SSD and Faster R-CNN, which are 88.09% and 86.57% separately. To perform bridge concrete crack quantification, several image processing techniques are applied. The process begins with the cropping of the identified cracks obtained by YOLOX_s followed by binarization using Otsu's method. Subsequently, the Zhang-Suen thinning algorithm is applied to extract the crack skeleton, while the Canny edge detection algorithm outlines the crack boundaries. Finally, a pixel accumulation-based method is implemented to calculate the crack dimensions. The findings indicate that the proposed method for measuring crack length and the maximum width achieves high accuracy levels of 96.6% and 95.86%, respectively.