Threshold Segmentation Bridge Crack Detection Algorithm Based on Deep Learning

Abstract

A crack segmentation and width measurement method based on deep learning and Zernike orthogonal moments was proposed. In response to the shortcomings of poor applicability of traditional methods, deep learning algorithms are combined with traditional algorithms based on digital image processing. Based on the qualitative classification ability of deep learning, the detection of cracks in images is divided into three levels: "judgment of presence", "automatic sketching", and "width measurement". Adopting multiple scales of deep learning to narrow the scope of cracks, and on the basis of traditional methods for preliminary segmentation of cracks, deep learning is used again to screen the preliminary segmented cracks, greatly improving the accuracy of crack segmentation and noise resistance in complex environments. In terms of width measurement, a method based on Zernike orthogonal moments for measuring the width of small cracks within 5 pixels is proposed to address the shortcomings of traditional "few pixels" methods, which have large measurement errors for small cracks within 5 pixels. The method directly uses the grayscale information of cracks to calculate the width of cracks and improves the accuracy of width measurement for small cracks in images.