Computer vision-based crack width identification using F-CNN model and pixel nonlinear calibration

Abstract

To improve the accuracy of computer vision-based bridge crack-width identification, two factors were investigated in this study. First, a fully convolutional neural network combined with a U-Net architecture was used to extract crack pixels and a crack midline (i.e., crack skeleton). A database including 100 images with 572 × 572 pixels labelled for cracks was developed. The results revealed a mean of 84.4% of the average mean intersection over union of the U-Net. In addition, a crack-width direction identification method based on the slope of the crack skeleton was proposed, and the accurate extraction of the nonuniform width parameter along the crack was realised. Second, to obtain a mapping from the width pixels to width physical dimensions, a pixel calibration experiment was conducted. Finally, a nonlinear regression model of the distance, focal length, and actual pixel size was developed to overcome the light distortion of the camera. Ultimately, the combination of the two factors completed a high-precision extraction of the entire process of crack-width identification and achieved a 0.01-mm precision under a 96% guarantee rate.