Research And Improvement Of Image Super-resolution Reconstruction Based On Convolutional Neural Network And Generative Adversarial Network

Image Super Resolution Reconstruction(Super Resolution,SR)is a classic application of computer vision,which refers to the reconstruction of low resolution images into high resolution images through software or hardware methods.Image super resolution reconstruction can be divided into two categories: high-resolution images reconstructed from multiple low-resolution images and high-resolution images reconstructed from a single low-resolution image.This thesis mainly studies the reconstruction method based on single image super-resolution.Image super resolution reconstruction has good application value in the fields of monitoring equipment,satellite image remote sensing,digital high-definition,microscopic imaging,video coding communication,video recovery and medical imaging.Due to the limitations of hardware technologies such as optical instruments,some image blurs cannot meet specific requirements,and software methods such as algorithms need to be used to further improve image resolution.Although the image super-resolution reconstruction based on convolutional neural networks and the image super-resolution reconstruction based on generative adversarial networks have solved the problem of traditional image resolution to some extent,there are still improvements in algorithms space.This thesis deeply researches the deep learning convolutional neural network and generative adversarial network algorithms.According to some problems of the two algorithms,a method of improving the two algorithms is proposed,which effectively improves the super resolution of the image,and the reconstructed image is more clearer and the details of the restored image are more abundant.The main research contents of this thesis are divided into the following points:This thesis proposes an improved method for image size reduction and image blurring after image super-resolution algorithm reconstruction based on convolutional neural network.The depth of the network layer is increased in the model,and the residual network is added.The improvement measures can not only extract more image details,but also accelerate the model training.The experimental results show that the improved model based on the convolutional neural network image super-resolution reconstruction algorithm proposed in this thesis reconstructs the image more clearly,restores more image details,and on the two image quality evaluation indexes,the peak signal to noise ratio(PSNR)and structural similarity index measur(SSIM)have been improved to a certain extent.This thesis introduces the relative generative adversarial network to solve the problems of the generative adversarial network itself to solve the problem that the Nash equilibrium point of the generative adversarial network is difficult to find,the model is difficult to converge and train.The image super-resolution model based on the generated adversarial network has a good effect on the reconstruction of images magnified four times or more,and can recover more high-frequency information.However,due to the complex network structure of the model,the difficulty of model training,and the high requirements on the experimental environment,especially the hardware conditions,this thesis uses a convolution-deconvolution method to modify the generator part of the algorithm model to make the generator The network becomes a completely symmetric network,while reducing the number of network layers.The relative generation of the adversarial network is used to modify the loss function of the discriminant network.The experimental results show that improved model based on the generative adversarial networks image super-resolution reconstruction algorithm proposed in this thesis reconstructs the image reaches the visual effect of the original model,and the image evaluation indexes of the peak signal to noise ratio(PSNR)and structural similarity index measur(SSIM)have been improved to a certain extent.