Research on a Crack Extraction Algorithm of Bridge Deck of Simple Supported Girder

The crack detection of the simple support beam bridge before leaving the factory is an extremely critical step in its production and casting process, which is directly related to whether the bridge can be serviced normally. In this paper, an improved bridge deck crack extraction algorithm based on ResNet and Fcn convolutional neural networks is proposed. Firstly, a 1-3-1 mode residual network is constructed based on ResNet network to extract the features of the input crack pictures. Then, the size restoration work is carried out by the deconvolution module, and a multi-scale information fusion module is added to increase the receptive field to ensure the transmission of detailed information. Experiments show that the average pixel accuracy of the proposed algorithm is 0.84 and the average intersection-union ratio is 0.81, which is better than other comparison algorithms, and can effectively complete the task of crack detection of the bridge deck of simple support girder.

I. Introduction

In recent years, deep learning has made significant breakthroughs in the field of image recognition with the introduction of various neural networks, and more and more scholars have started to study crack detection techniques based on various convolutional neural networks. in 2015, makantasis [1] constructed cnn convolutional networks to achieve the detection of tunnel cracks. in 2017, zhao [2] used the zhalexnet model as the basis for developing a detection method for scratches and cracks on the surface of steel materials. In 2019, Song et al. [3] employed a lightweight neural network model called SegNet to perform segmentation extraction of crack contours with a maximum IoU of 0.782. in 2020, Chen, FC [4] proposed a rotation-invariant fully convolutional neural network (Fcn) called ARF-Crack for crack segmentation. Choi et a1. [5] constructed SDDNet for crack extraction. The model consists of a standard convolution, a densely connected separable convolution module, a modified void space pyramidal pooling module and a decoder module. Nowadays, various methods of deep learning have become mature in the recognition of cracks, but there is still a deficiency in the extraction of crack contours, which cannot extract the edge detail information of cracks well.