Research And Improvement Of Image Super-resolution Based On Generative Adversarial Networks

Image super-resolution technology refers to the technology of reconstructing a high-resolution image by processing low-resolution images complementary to one or more pieces of information and is widely used in medical imaging,video surveillance,remote sensing imaging and the like field.The learning-based image super-resolution method implicitly learns the mapping between high-resolution and low-resolution images and then super-resolutions the input image by learning the mapping function.In recent years,with the introduction of deep learning-based methods into the field of image super-resolution,the learning-based approach presents significant advantages in super-resolution algorithms.Generative adversarial networks have received widespread attention since being proposed and has been applied to image segmentation,image style conversion,image super-resolution and other fields.Generative adversarial networks consist of the generator model and discriminator model,the two games with each other until it reaches Nash equilibrium.Based on the method of generating confrontation mesh,it can restore the texture information and the grain details to the greatly downsampled image.However,the super-resolution method based on generative adversarial networks can only deal with a kind of magnification and lack of universality.When dealing with the complicated scenes of real scene,the model's ability of representation is not high.Moreover,generating anti-network training is very unstable,Crash occurs,seriously affecting the quality of the generated image.This paper introduces the basic flow and principle of image super-resolution algorithm based on generative adversarial networks.In this paper,the following improvements have been made to the generator model: The multi-cascade structure is used to magnify images step by step so that the model can generate multiple magnification images at the same time,and at the same time ensure higher quality images obtained at larger magnifications;By using recursive learning and residual learning to improve the Res Net model and improve the model's expressive ability,the quality of the generated images is significantly improved.An Expand-Squeeze method is proposed to generate a picture.The basic idea is to extend the model lastly A layer of convolutional dimensions to get more contextual information,and then generate the image using a 1x1 convolution kernel to effectively reduce the checkerboard effect.In this paper,we improved the discriminator network structure,introduced the distance to measure the similarity between the generated image and the real image,solved the collapse problem when generating the anti-network training,improved the training stability of the model,and then improved the quality of the generated image to some extent.This paper uses Set5,Set14,BSD100 three public data sets for testing.Experimental results verify that the improved method of generative adversarial networks in this paper can effectively improve the quality of the generated image and effectively improve the training stability of the model.At a magnification of 8 times,the improved model is based on Set5,Set14 andBSD100 data sets PSNR values reached 26.22 dB,24.58 dB and 24.61 dB respectively.