Image Super-resolution Using A Generative Adversarial Network

As one of the main media for information dissemination,images are widely used in various scenes.In many field,such as video surveillance,satellite remote sensing,and medical imaging,low-resolution images are no long sufficient for applications.Although the traditional super-resolution method has achieved a lot of results in recent years,it can basically recover high-resolution images,but problems such as image blur and loss of detail still exist,resulting in the missing key information of the image.The image super-resolution algorithm based on deep learning can solve the problems of large computational complexity,low reconstruction performance,and low image quality.However,there are problems exist in cast of blurred and distorted image,such as such as insufficient network training,unstable models,gradient disappearance,etc.In order to solve the above problem,this paper combines the idea of convolutional neural network,and deeply studies the optimization method of generative adversarial network in the field of image super-resolution.The convolution neural network model with multi-feature fusion and the generative adversarial network model with residual compensation are proposed.The above models are applied to the image super-resolution field.The main research contents of this paper are summarized as follows:(1)Analyze the typical image super-resolution algorithm,summary the basic principle of convolutional neural network in detail,study the model structure of generative adversarial network deeply and confirm the advantage and disadvantage of the convolutional neural network and the generative adversarial network in the image super-resolution field in theoretically.(2)Considered of the problems existing in the current image super-resolution algorithm based on convolution neural network,such as single feature extraction,insufficient network training and low image quality,an image super-resolution algorithm based on multi-feature fusion convolution neural network model is proposed.The algorithm improves image information integrity by combining high-frequency detail features and low-frequency structural features.And through the sub-pixel convolution layer to optimize image reconstruction performance,reduce model parameters,and accelerate network training.(3)Considered of the problems existing in the current image super-resolution algorithm based generative adversarial network,such as the unstable network training and the pool visual effect of the generated image,the generative adversarial network is used as the basic framework and the result of Bicubic interpolation is compensated as residual to the generative model in this paper,which completes the detailed information of the image and enhances the visual effect of the reconstructed image at a higher magnification.In addition,the Wasserstein distance optimization discriminative model is used to improve the stability of the algorithm training.This paper uses the common data sets Set14,Set5 and BSD100 for testing.The experimental results show that the improved generative adversarial network model and convolutional neural network model can effectively improve the quality of reconstructed images and enhance the stability of network training.The peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)of reconstructed images are both improved and the visual effect of reconstructed image is better than other algorithms.