Research On Blind Image Super-resolution Algorithm Based On Generative Adversarial Network

Image resolution is an important factor to measure image quality.Compared with lowresolution images,high-resolution images have more detailed features and clearer edge information,but it is more difficult to obtain high-resolution images.In recent years,superresolution reconstruction technology has been widely used in many fields such as medical imaging,public safety and satellite remote sensing.Therefore,super-resolution reconstruction technology has always been a research hotspot in the field of image processing.Most current reconstruction techniques are non blind algorithm models,and the target image is a low resolution image from existing datasets,which cannot meet the image reconstruction needs in real scenes,Compared with other technologies,blind image super-resolution technology can use existing data to learn and accurately reconstruct low-resolution images with unknown degenerated kernels,which has high research value and application prospects.With the emergence of convolutional neural networks,image super-resolution reconstruction technology has developed rapidly.This paper mainly uses generative adversarial networks to build a blind image super-resolution algorithm framework,and conducts experimental demonstrations on its reconstruction performance.The main research contents are as follows:(1)A multi-scale blind image super-resolution model is built to solve the problem of poor recovery of reconstructed image detail features.Specifically,combined with the idea of generative adversarial network,an attention mechanism and texture enhancement module are designed in the generative model,and the image features are weighted nonlinearly through parallel connections of channel attention,spatial attention and self-attention,so that It is easier to extract image features.The texture enhancement module is used to mine the texture features of different scales of the input image,enhance the texture details,and improve the ability of the generative model to learn features.Combined with the discriminant model of UNet++,the information is fed back to the generative model pixel by pixel,which can retain all prior information and improve Gradient,so that the model can use the perfect feature information extracted above during the training process.(2)In order to seek the influence of the size of image receptive field on model learning features,a model based on receptive field optimization is proposed.A dense residual module is designed to connect shallow features and deep feature residuals through global feature fusion to increase the flow of information and gradients while ensuring that feature information is not lost during the circulation process.The receptive field optimization module is used to expand the effective receptive field of the input image without affecting the resolution by using hole convolution,and the channel information of the output is mixed by point-by-point convolution,which is more convenient for the model to learn features.The training data set is used to train the generative model first,and then train together with the discriminative model.Experiments prove that the model improves the performance of blind image super-resolution reconstruction through the combination of modules.(3)Conduct a large number of comparison experiments and ablation experiments on the model proposed in this paper,and select the existing test data set and real low-resolution images to test the performance of the model.The experimental results show that compared with the existing advanced methods,the proposed method has better perceptual quality of the reconstructed high-resolution image,and the detailed features and edge information of the image have been well restored,the proposed blind image super-resolution reconstruction model performs well.