Crack Recognition And Detection Based On Deep Learning

With the large-scale construction of roads and bridges in China,there are more and more concrete structures,and crack damage detection for concrete structures is an urgent problem to be solved.The traditional crack detection of concrete structure includes manual detection,machine vision,etc.This type of detection method has a general effect in extracting concrete crack characteristics,and its recognition accuracy does not reach the expected effect.With the development of computer hardware and equipment,deep learning has shined.In 2016,AlphaGo defeated the world champion of Go,making the influence of deep learning rapidly expanding from academia to industry.Deep learning is rapidly applied in fields such as image recognition.Because neural networks have their unique advantages,they are far superior to traditional machine vision in extracting image features.This article applies deep learning to the classification and detection of concrete cracks,mainly in the following three parts.(1)Crack data collection.Collected a bunch of crack image data through cameras,mobile phones and other equipment,manually performed cutting,classification and other operations to construct a training data set;and rotated,scaled,sheared,translated,clipped noise and other series of image data operating.(2)Construction of crack classification model.Three classic networks AlexNet,VGG13,ResNet18 are used.In order to prevent overfitting,regularization,dropout,image enhancement,batch normalization and other measures are adopted;through experimental verification,ResNet18 is better than the other two networks,and each index is Better than AlexNet,VGG13;Finally,the actual pictures were tested,and the high-resolution pictures were classified using the sliding window method.(3)Construction of crack detection model.Annotate the crack pictures,construct the training data set of the crack model;select the latest model YOLO v3 in the YOLO series,and take a series of measures to prevent overfitting;the experimental results show that the model finally converged,test several pictures The crack detection effect is very good;the crack video is collected and detected by the trained model,which can accurately detect the crack area.