The Research Of Pavement Crack Image Super Resolution Reconstruction Based On Generative Adversarial Network

At present,because the collection of pavement crack images is affected by factors such as imaging equipment and the external environment,the acquisition of high-quality crack image data cannot be guaranteed.If the crack recognition or pavement performance analysis is directly performed on the pavement crack images,the results usually have large errors.Therefore,it is necessary to reconstruct the collectedpavement crack images with super-resolution.Judging from the current research status of image super-resolution reconstruction,the performance of the reconstructed images in texture details is still far from the original high-resolution images.And no super-resolution method is designed for pavement crack images.For the existing problems,this paper integrates the attention mechanism that can improve the image detail performance into the generative adversarial network.The research and design of super-resolution reconstruction method for pavement crack images were carried out.Finally,based on the assessment method of pavement crack images,a comparative experiment was carried out.The main research work is as follows:First of all,three typical super-resolution reconstruction methods based on convolutional neural networks,deep recursive convolutional neural networks,and generative adversarial networks are studied.The training and testing of image super-resolution reconstruction models were carried out using the collected pavement crack image data set.In order to make the models achieve the best results,different network parameters are also set,and multiple results are trained for tuning.Based on the best effect models,the advantages and disadvantages of different methods are summarized through a comparative analysis of the experimental results.It provides a clear direction and target for the research and improvement of super-resolution reconstruction methods.Secondly,attention mechanism is introduced into the generative adversarial network.The structure of the generator feature extraction module in the original generation adversarial network and the design of the discriminator loss function are improved.The generator is divided into three parts: shallow feature extraction network,nonlinear mappingnetwork and upsampling network.The designed attention recursive network replaces the convolutional layer cascade structure in the original nonlinear mapping network.Among them,the attention recursive network also introduces gate recurrent units and improved dense residual blocks.The improved generative adversarial network has better visual perception ability and deep feature extraction ability.It also reduces the scale of network parameters and the complexity of training.Finally,it is verified through experimental comparison of super-resolution reconstruction of pavement crack images.Compared with the original method,the reconstructed image is clearer and the texture details are improved.The PSNR and SSIM values of the improved method on the pavement crack image dataset reached 29.21 d B and 0.854,respectively.Finally,a comprehensive assessment method for super-resolution reconstruction of pavement crack images is designed.The original image super-resolution reconstruction assessment method can only give a quantitative assessment value on the image quality,but cannot make a qualitative assessment on the images of a specific application scene.Therefore,after the pavement crack images reconstruction,the images are also subjected to crack detection and crack segmentation processing.Finally,the experimental results show that the reconstructed crack images have improved the accuracy of crack processing.