Research On Mars Image Super-resolution Method Based On Generative Adversarial Network

With the successful launch of the "Tianwen-1" Mars probe,our country has begun exploration of Mars.High-resolution Mars images are of great significance for studying the landform features of Mars and analyzing the climate on the surface of Mars.Nowadays,the main image super-resolution algorithm is a deep learning method,which is better than traditional methods.However,in the current image super-resolution method,the low-resolution(LR)image is usually obtained from the high-resolution(HR)image through an ‘ideal’ downsampling method.The LR-HR paired obtained by using such a method is tested on an ideal dataset and the model obtained after training has produced impressive results.But once the real image is used for testing,it produces relatively poor results,such as some artifacts and at the same time,there are still shortcomings in super-resolution image details.In order to solve the above problems,we proposed a new Mars image degradation framework and Mars image super-resolution network.Firstly,we propose a new image degradation framework for real Mars images.Utilizing the powerful generalization ability of Generative Adversarial Network(GAN)to learn the degradation process of real images and simulate the noise distribution of real images.The more ‘realistic’ low-resolution images obtained by using the new image degradation framework are sent to the super-resolution network and obtain a better super-resolution result.Secondly,we use the principle of GAN Inversion to study the super-resolution method of Mars image and design a new image super-resolution network,propose a‘Encoder-Latent Bank-Decoder’ architecture.First,an ‘Encoder’ to perform multi-level feature extraction to generate more detailed image information.Then,use the improved Style GAN network as a generative latent bank to capture the prior information of the image.Last,a ‘Decoder’ is used to fuse the features in the encoder and the generative latent bank.Such a network structure only needs a single forward pass to generate the upscaled image.Finally,in order to meet the task requirements of image super-resolution,Three improvements have been made to the Style GAN network structure.First,instead of taking one single latent vector as the input,each block of the generator takes a different latent vector.Second,the multi-resolution features extracted by the encoder are also added to the generative latent bank,and the convolutional layer is used for feature fusion to further capture the features of the LR image.Third,instead of directly outputting the result of the generator,it outputs the feature vector of each augmented block of the feature.Such an improvement can reduce the pressure of the generator to generate image details,at the same time,it can better integrate the output features from the generative latent bank and the encoder.