Research On Bridge Crack Detection Method Based On Convolutional Neural Network

As a common infrastructure,bridges play a pivotal role in daily life.However,the condition of the bridge will deteriorate due to environmental and loading effects,and in severe cases it will affect the safety of pedestrians and vehicles.Therefore,it is necessary to regularly check the current condition of the bridge and repair the damaged area in time.The traditional bridge maintenance methods are mostly based on manual inspection,which not only takes a long time,but also is easy to miss and misdetect,and the detection efficiency is low.The bridge crack detection technology based on digital image processing has the advantages of fast detection speed and convenient detection,and has gradually replaced the manual detection method.However,bridge crack images often contain complex background information,and digital image processing methods are susceptible to noise interference,resulting in reduced detection accuracy.Aiming at the above problems,this paper studies the crack detection method based on convolutional neural network,and designs the RES-ASPP(Residual Atrous Spatial Pyramid Pooling)model based on Res Net50 migration learning pre-training and an efficient,high-precision,and no pre-trained MSA(multi-scale atrous convolution model)model.The main tasks completed in this paper include:1.The research and analysis of the existing crack detection algorithms based on deep learning is investigated and analyzed.The bridge crack detection algorithm based on deep learning is studied and proposed.The algorithm flow includes input image,preprocessing,data set calibration,model training,crack detection and output result.2.The artificial amplification scheme of the crack data set is designed,the image cutting and image enhancement of the existing limited mass bridge crack data set is produced,and the data set containing 6069 bridge cracks and background images is generated,which can be used for the subsequent bridge crack detection.3.The RES-ASPP model for bridge crack detection based on Res Net50 is studied,the model is pre-trained on the Image Net data set,and the ASPP module can extract multi-scale feature information is added according to the task requirement of bridge crack detection.Performance tests are performed on the crack detection dataset.The experimental results show that the model has achieved good detection results.4.An efficient and high-precision MSA model is proposed,which takes the advantage of atrous convolution,Dep-ASPP(Depthwise Separable Atrous Spatial Pyramid Pooling)module and depthwise separable convolution.Accuracy of crack detection of 96.37% was achieved without pre-training.At the same time,the model has higher computational efficiency and fewer model parameters than the traditional classification models.The floating-point operations(FLOPs)consumed are lower than the traditional classification models Res Net18,VGG16,etc.,and it takes only 286 minutes to train 300 epochs.