Bridge Crack Detection System Based On Convolution Neural Network

In recent years,with the vigorous development of computer vision and machine learning,many domestic and foreign experts are devoted to applying the method of digital image processing based on machine learning to the bridge Crack detection field.At the same time,unmanned aerial vehicle technology has also been widely used in industry.Using a drone equipped with a camera to capture a picture of a bridge crack has become an important means of replacing the artificial detection of cracks.But when using UAV to take pictures of bridge cracks,the amount of pictures taken is very huge.If we use traditional digital image processing methods to identify whether there are cracks in all images,it will take a long time and can only recognize the shape Simple and obvious crack pictures,once the crack shape is complex,traditional digital image processing methods can not get the desired effect.Based on this,this paper applies the convolution neural network to the detection and classification of bridge cracks innovatively,and can classify the images collected by the UAV efficiently and screen out a part of the pictures with real cracks,then screened images for subsequent crack length and width measurement.In this paper,a convolution neural network is designed,which can directly input the crack images captured by UAV,avoid complex pre-processing and improve the recognition efficiency.However,the actual crack shapes are largely random,and unlike other bridge workpieces,the fractured images do not have obvious spatial geometric features,which makes it difficult for the input layer of the convolutional neural network to extract the features of the original fractured images.If an image without preconditioning is directly input to the convolutional neural network,the neural network can not converge and the training can not be performed.In order to make the training of convolutional neural network more effective,this paper innovatively introduces two-dimensional information entropy into the fracture image evaluation.Based on this,a corresponding experiment has been carried out and a preprocessing method has been found to significantly enhance the structural information of the fracture image.By using the preprocessing method,the information entropy of the crack picture can be effectively enhanced without changing the information entropy of the non-crack image as much as possible.So the structure information extracted by the convolutional neural network can be increased to construct a feasible convolution neural network.After a large number of parameter adjustment experiments were carried out,the parameters of the actual convolution neural network were obtained,and the filter window size selection experiment and decision threshold setting experiment were carried out.Finally,we obtained a complete Which is suitable for the smart identification scheme of crack identification.After the final real image recognition experiments,and the performance of the convolutional neural network in this paper is verified.Through the actual identification experiment,we can conclude that the convolutional neural network in this paper has good practicability.In addition,this paper also explores the use of traditional image processing methods to measure the length and width of the crack information.Finally,a complete bridge crack detection system software was developed.This software combines the rapid identification of bridge cracks based on convolution neural network and the length-width measurement algorithm based on the traditional image processing method.Realized the automatic full link from the UAV acquisition of the picture to calculate the length and width of the bridge crack.