Research And Application Of Single Image Super-Resolution Algorithm Based On Deep Learning

The image processing problem has been widely studied and applied in the past few decades,and the fierceness in the field of deep learning has provided new research tools for this problem.Models based on deep convolutional neural networks have achieved great success in the field of image processing,and it has become a research hotspot in the field of image processing in recent years.This paper is mainly based on the research on the practical problems of generative adversarial network development and algorithm improvement in deep learning network models.The main work done is as follows:(1)In the single-image super-resolution algorithm design,most of the models based on the deep convolutional neural network cannot make full use of the layered features of the original low-resolution image,resulting in chaotic image details and low credibility.In order to improve the high-frequency detail quality of reconstructed super-resolution images,a super-resolution method of residual dense generation adversarial network(RDGAN)is proposed.This method uses generative adversarial networks as the main model structure and residual dense blocks as the basic building blocks of the generator.This allows the network to focus more on the layered features of low-resolution images and then use convolution layer extracts all the hierarchical information.At the same time,the perceptual loss is used as the loss function of the model,so that the generated image can have better detail texture.It can be found in the evaluation of the model effect that compared with the best models at present,this model can make the super-resolution images have higher PSNR and SSIM values.(2)In the study of removing raindrops from a single image,raindrops attached to glass windows or camera lenses can severely degrade image quality,so use a generation adversarial network to convert degraded images with raindrops into clean images.The article introduces the visual attention mechanism into the generation network,proposes a new residual U-Nets to process the attention map,and uses a discriminative network to discriminate the authenticity of the image,while using a new perceptual loss as a loss function,so that the network can be more Pay attention to the raindrop area and its surroundings.Experiments show that this method can get higher SSIM value and improve the recognition quality in image recognition compared with the best methods at present.(3)The scarcity and confidentiality of medical image data sets cause serious imbalances in medical images.In order to solve this problem,a high-quality color fundus image data set is obtained,and a generative adversarial network is proposed to generate fundus images.By comparing the Pix2pix and Cycle-GAN models with different structures,it is shown that using the multi-feature fusion input of the optic cup,optic disc,and blood vessels is significantly better than using only single blood vessel inputs.The experiments were performed on two public fundus image datasets(DRIVE and DRISHTI-GS),and the results show that the multi-feature fusion input method can generate color fundus images with brighter discs and fuller details.