| With the continuous improvement of road construction,the biggest problem faced is road maintenance,and cracks are one of the most common reasons for road problems.Accurate detection of cracks can prevent potential problems and reduce the large cost and human resources in road repair,playing a crucial role in road construction.Traditional detection methods rely on manual recognition and wireless sensors,which have low efficiency and high cost,and are not conducive to improving detection accuracy.This article proposes crack detection based on an improved U-net neural network,utilizing the full "U" structure of the U-net network,and combining attention mechanism and residual units of the Resnet network to improve the network.A highprecision detection network is designed,with the following specific contents:1.By learning real images through generative adversarial networks,random noise is transformed into regular vectors,and the dataset is continuously adjusted through loss feedback to expand and reduce overfitting.To address the issue of blurring generated images,a self attention mechanism is introduced to focus the generator on generating details.2.In order to fully utilize the true information of the target,it is necessary to perform preprocessing operations on the image,including denoising,graying,and image enhancement.Multiple filtering methods are used to denoise the image,and the RGB three channel image is converted to grayscale using the weighted average method.Finally,the original image is further improved through image enhancement.This algorithm can effectively reduce the impact of redundant information in the image on the target,improve the detection ability of the target,and make the detection of the target simpler,thereby enhancing the credibility of the detection.3.Based on the improvement of U-net network structure,the traditional convolution is replaced by cavity convolution to increase the Receptive field and improve the extraction accuracy of crack feature information.At the same time,in order to prevent excessive learning of redundant information in the network layer,Dropout is added layer by layer,and LRN modules are added in the interlayer to improve the generalization performance of the model.The improved network model was applied to a self-made crack dataset,and the crack detection performance of the existing network and the improved network was compared.The results were analyzed.The experimental results showed that the loss value during the training process of the improved U-net network reached 0.0015,with an accuracy of 0.98.Compared to the traditional U-net network,the loss value in crack detection decreased by 0.001,and the accuracy increased by 0.12,Compared to the original U-net network,the recall rate and intersection ratio have increased by 0.0193 and 0.1573,respectively,showing good crack detection performance.4.In response to the poor performance of the improved U-net network model in crack detection and the problem of edge information loss,further improvements are made to the improved network model by adding attention mechanisms at the skip connections between downsampling and upsampling to avoid the impact of image noise on the model,improve the extraction of key crack features,and introduce residual units into the convolutional layer,And the improved model was experimented on a single background image and a complex background image,and the results were analyzed.Compared with the improved U-net network model in the previous chapter,the recall and intersection ratio were improved by 0.0542 and 0.0176,respectively,verifying the effectiveness of the proposed model in crack detection. |
