Design And Implementation Of Image Super-Resolution Algorithm Based On Generative Adversarial Nets

Image super-resolution is a classic problem in the field of image processing,and more and more researchers are conducting related research.This thesis focuses on the use of a single low-resolution image,using deep neural networks combined with software algorithms to calculate the correlation between low-resolution images and high-resolution images,and finally can achieve a corresponding high resolution for any image.Rate image.This approach is more flexible and less expensive than upgrading the image spatial resolution by hardware upgrades.In this thesis,two kinds of deep learning methods based on deep convolutional neural networks and generating anti-networks are studied.The optimization methods are proposed respectively.In the optimization of the former,the network performance is further improved by deepening the network depth,and the residual network is introduced to avoid the problem of degradation caused by the network being too deep,and the network model is modified to move the super-resolution step to the network.This makes it possible to map low-resolution features to high-resolution outputs through training and learning,so that convolution operations are not required at higher resolutions,thereby reducing computational complexity.Finally,it is verified by experiments that the optimization algorithm has better superresolution effect.In the latter optimization,the network model is designed to combine the deep convolutional neural network to generate the confrontation network,which makes it more suitable for solving the problem of image super-resolution,which greatly improves the speed and stability of training.The traditional algorithm verifies the effectiveness of the proposed method.At the same time,the experimental verification shows that if the image type of the training set is consistent with the image type to be reconstructed,better results can be obtained.In the experiment,the CelebA face dataset image is used as the training set.It is confirmed that the same type of image is trained and predicted,and the super-resolution effect is better.