Research Of Generative Single Image Super Resolution Methods Based On Image Distributions

Thanks to the development of multimedia technology on computers,digital image processing technology and digital image display technology,people’s demand for displaying high-resolution images is gradually increasing.However,subject to external conditions and equipment limitations,it is very difficult to directly obtain highresolution images.Therefore,image super-resolution technology came into being.Super Resolution(SR)technology is an image processing technology that reconstructs a corresponding high-resolution image from one or more blurred low-resolution images.With the continuous development of deep learning technology and the significant improvement of computer computing power,deep learning techniques have also been widely used in super-resolution imaging problems.However,existing deep learning-based methods either choose to optimize the super-resolution process from pixel information,or choose to optimize image information from a perceptual perspective.These optimization methods all lack the utilization of image distribution information.Taking the commonly seen single-frame image super-resolution task in image superresolution tasks as the research object,this article takes full analysis and utilization of distribution information as the starting point,and aims to achieve better reconstruction results and distribution transfer in image super-resolution.Three image super-resolution models were designed.The main work contents are as follows:Image Super-resolution based on dual discriminator Generative Adversarial Networks Generative Adversarial Networks(GANs)have become a popular choice for image super-resolution tasks due to their ability to generate high-quality images.However,the existing GAN-based super-resolution methods suffer from the problem of producing blurry and artifact-filled images in the output.In order to address this issue,we propose a novel dual discriminator GAN(DWGAN)structure for single-frame image super-resolution.The DWGAN model includes two discriminators: the original discriminator used to judge the image output and an additional discriminator to constrain the hidden space distribution.By incorporating the distribution information,the proposed model can better capture the underlying image features and produce more accurate and visually appealing super-resolved images.Experimental results on three publicly available datasets demonstrate the superiority of the proposed DWGAN model over the existing state-of-the-art super-resolution methods in terms of both quantitative and qualitative measures.Specifically,the DWGAN model achieves better performance in terms of peak signal-to-noise ratio(PSNR),structural similarity index(SSIM),and visual quality.Furthermore,we conduct a series of ablation experiments to demonstrate the effectiveness of the proposed innovative modules in the DWGAN model.The results show that the proposed modules effectively improve the performance of the model and contribute to the superior super-resolution results.In summary,the proposed DWGAN model provides an effective solution for single frame image super-resolution by leveraging the distribution information in the image.The model can produce high-quality super-resolved images with better visual quality and overcome the common boundary problems in super-resolution.The experimental results validate the effectiveness of the proposed approach and demonstrate its potential for practical applications in various fields such as medical imaging,surveillance systems,and satellite imaging.Image Super-resolution based on Wasserstein Auto Encoders While the generative adversarial network-based method can use distribution information to reconstruct images,its training process is often unstable and may lead to mode collapse.In order to address these issues,we propose a new deep learning method based on the Wasserstein autoencoder for image super-resolution tasks.We add a penalty term to the training process that incorporates the distribution of the output image and the distribution of the real image,improving the network’s performance on super-resolution tasks.We evaluate our method on four publicly available datasets and show that it outperforms other existing image super-resolution methods.Additionally,our method is effective in solving the common issues of complex texture images,over-smoothing,and loss of detail during the image reconstruction process.This paper also conducts extensive ablation experiments to demonstrate the effectiveness and versatility of our innovative modules in the proposed network.By utilizing the Wasserstein autoencoder-based method,we can achieve better image superresolution results while avoiding the instability and mode collapse issues encountered by the traditional generative adversarial network-based methods.The penalty term we introduce to the training process is a novel approach to incorporating distribution information and contributes to our network’s superior performance.Our experiments show that the proposed method can significantly improve the quality of the reconstructed images,producing visually pleasing results with more accurate and detailed features.Distribution transformable image super-resolution methods based on Wasserstein Auto Encoders We propose a novel distribution-transferable image super-resolution method based on Wasserstein autoencoders,inspired by reference-based image superresolution and style transfer tasks.Our method introduces an additional image as input to provide distribution information for the output,and transfers the distribution of the reference image to the output image through a distribution transfer module.We conducted experiments on four public datasets and demonstrated that our network outperforms other single-image super-resolution networks when the distribution transfer module is not considered.Additionally,our network dynamically adjusts the transfer degree of the distribution as the transfer coefficient α is adjusted.Furthermore,we conducted ablation experiments to prove the effectiveness and necessity of each module in the network structure.The experimental results show that our proposed method achieves state-of-the-art performance and is capable of solving common problems in image super-resolution.Focusing on the task of single-image super-resolution,this paper proposes a series of algorithms starting from modeling and comparing image probability distributions.The proposed methods include a two-stage generative adversarial network(GAN)based image super-resolution algorithm,a Wasserstein autoencoder-based image superresolution method,and a distribution-transferable image super-resolution method based on Wasserstein autoencoder.The two-stage GAN-based method achieves the superresolution task by constraining the distribution in the latent space.The distributiontransferable image super-resolution method based on Wasserstein autoencoder utilizes the difference between the distribution of the real image and the output image.This method can achieve distribution transfer on images with different styles.Experimental results demonstrate that all three networks can achieve the desired super-resolution performance with limited training data.