Research And Implementation Of Bridge Crack Detection Technology Based On Deep Learning

With the development of transportation infrastructure in China,bridge plays an increasingly important role in China’s economic and social life.Therefore,it is very important to test the health degree of bridge structure periodically.At present,concrete structure is the main bridge structure in China.Due to external factors and internal stress,the concrete structure will inevitably crack.The occurrence of cracks will reduce the bearing capacity and durability of the bridge structure,thus leading to many bridge diseases.Therefore,bridge crack detection has become an essential part of the bridge safety detection process.In the past,bridge crack detection work mainly relies on manual work,which has the disadvantages of strong subjectivity,high labor density and high detection cost.To solve this problem,this thesis proposes a lightweight,high-precision and automatic bridge crack detection technology based on deep learning algorithm.The main work of this thesis is as follows:(1)Aiming at the problem that the collection process of bridge fracture data sets is cumbersome and the number of data sets collected in a short time is small,which is not enough to train the deep learning model completely,a data set expansion method combining supervised data enhancement and unsupervised data enhancement is proposed.A mixture of image rotation,image cropping,noised image,contrast enhancement and deep convolution generative adversarial network is used to expand the bridge fracture data set.After the expansion of this method,a total of 40,000 bridge fracture images are obtained,among which 20,000 are positive samples and 20,000 are negative samples.And annotate the expanded data set to provide data support for later model training.(2)In view of the complex and diverse texture and irregular distribution of bridge crack images,the prior box selection strategy of the deep learning target detection model cannot meet the size requirements of all the crack images,and the missing detection phenomenon of detection results is serious,a gridded bridge crack detection method based on depth classification is proposed.Firstly,the image of bridge cracks is divided into equal grids,and the location distribution of bridge cracks in the image is normalized.Secondly,different depth classification models are compared and analyzed.According to the experimental results,the most suitable depth classification model is selected to replace the target detection model to reduce the complexity of the algorithm.Finally,according to the recognition results of depth classification model,the grid area identified as fracture target is automatically marked to achieve the purpose of bridge fracture location.After training with the expanded data set,the proposed method is compared with the detection results of current advanced target detection algorithms YOLOv4 and Faster-RCNN.Experimental results show that the proposed method is far superior to the latter in the accuracy,recall rate and F1 score of bridge crack detection task.(3)Aiming at the problems of large number of model parameters and high storage cost of the depth classification model VGG16 selected in the designed detection technology,the lightweight improvement of the model was carried out according to its parameter distribution characteristics.The penultimate two full connection layers of VGG16 model full connection module were replaced by global average pooling.Delete the last convolutional layer of each convolutional module;The channel pruning technique based on batch normalization layer is used to prune redundant convolutional channels.The experimental comparison of the model before and after the lightweight improvement shows that the number of model parameters is reduced from 275 MB to 0.34 MB under the condition that the detection accuracy of the improved model is approximately unchanged,which verifies the effectiveness of the lightweight improvement scheme designed in this thesis.Based on the deep learning algorithm,this thesis studies the bridge crack detection technology.Through data set preparation,the design of lattice bridge crack detection method based on depth classification,and the lightweight improvement of depth classification model,the design and implementation of lightweight grid bridge crack detection technology based on depth classification is finally completed.