Research On Super-resolution Method Of Train Detection Image Based On Deep Convolutional Neural Network

Super Resolution(SR)does not change the imaging equipment and imaging conditions,but improves the spatial resolution of the image through digital image processing.Deep convolutional neural networks have obvious advantages in processing computer vision tasks.Combining it with image super-resolution can accurately and efficiently improve the quality of image super-resolution reconstruction.In recent years,the railway transportation industry of China has developed vigorously,and train operation safety has become more and more important.The application of superresolution technology in the field of train images can reconstruct low-resolution blurred images into clear and delicate high-resolution images.And high-resolution images of the trains and their components are very important for the safe monitoring of the train’s operation.Research on the problem of insufficient use of feature information by deep convolutional neural networks and the inability of pixel loss-based networks to overcome the artifacts and smoothing of reconstructed images,and analyze the specific structure,network modules,loss functions,etc.of deep convolutional neural networks.Through the construction of information distillation and dense feature fusion module to enhance the flow efficiency of the network internal information;the introduction of an improved generative confrontation network to improve the reconstruction of image artifacts and over-smooth problems.The above two endto-end super-resolution network models are used to optimize the super-resolution reconstruction of train detection images,as follows:First of all,with the development and application of convolutional neural networks,the depth of the network is also increasing,and the number of parameters of the network has increased dramatically.However,due to the insufficient utilization of the feature information at all levels of the network,there is still room for improvement in the quality of image reconstruction.This thesis proposes an image super-resolution reconstruction network based on information distillation and dense feature fusion.This model uses a dense feature fusion structure combined with a residual network to fuse the deep and shallow features extracted by the information distillation module,so as to achieve the effect of less reconstruction time but higher performance.Second,to solve the problem that the existing super-resolution image reconstruction algorithm based on pixel loss has poor effect on high-frequency details such as texture,an image reconstruction algorithm based on improved super-resolution generative countermeasure network is proposed.First,in order to prevent artifacts,we remove all the batch normalization layers in the generator,and combine the multi-level residual network and dense connections,and use the residual set of residual dense blocks to improve the ability of the network to extract features.Then,the mean square error and the perceptual loss are combined as the loss function to guide the generator training,which not only preserves the high-frequency details of the image,but also avoids the appearance of artifacts.Finally,the last Sigmoid layer of the discriminator network is removed to better converge the training process,and the relative function is used to guide the training of the discriminator.The experimental results on the COCO data set and train data set show that this algorithm has improved peak signal-to-noise ratio(PSNR)and structural similarity(SSIM),and the average opinion index(MOI)and visual effect of this algorithm are far better than other algorithms.