Research On Single Image Super-Resolution Algorithm Based On Convolutional Neural Network

Image super-resolution refers to the process of recovering a corresponding highresolution image(HR)from a low-resolution image(LR).Video and images have become an important medium for modern people to contact the world,and image technology is also widely used in real life,such as medical images,supervision and security,short videos,etc.Since these applications have high requirements on image clarity,the cost of improving the resolution through hardware devices is high,and the ill-posedness of the problem needs to be better suited to the problem,so it is still challenging to give a reasonable solution.At present,software algorithm is a more economical and effective way to improve image quality.With the expansion of the application scope and its role as the cornerstone of computer vision,this topic has become an important part of the current computer vision field.For the task of image super-resolution,on the basis of convolutional neural network,this paper studies how to efficiently extract features,improve model performance,and further researches are conducted to solve the problem of image super-resolution with more interfering factors.The specific research content as follows:(1)An image super-resolution model based on frequency division combined with attention mechanism and multi-scale cascade residual is proposed.The network model consists of three parts,including image frequency division,feature extraction,and the final reconstruction module.First,the high and low frequencies are separated by mean filtering,and the details corresponding to the high frequency information are extracted through multi-scale cascade residual and attention mechanism;the image content corresponding to the low frequency is obtained through the channel random mixing operation to obtain the corresponding tensor elements for inter-channel communication;The final fusion stage reintegrates the high and low frequency to reconstruct the image.Through ablation experiments including and reference verification of some existing methods,it is proved that our method has better performance in terms of effect and performance.(2)A progressive single-image super-resolution algorithm based on generative adversarial network and attention mechanism is proposed.Given the noise of images in most real scenes and the lack of ground-truth labels,we extract images from images in a self-supervised manner.The reference information is obtained by itself,and the commonly used bicubic is replaced with an anisotropic Gaussian kernel for image degradation.During the training process,a noise directly obtained from the input data set via a sliding window is randomly selected and injected to simulate the real noise generation.Large-scale image super-resolution is achieved through progressive upsampling.When it comes to network model,the generator consists of Residual-in-Residual Dense Block(RRDB)and an improved channel and spatial attention series module(CSAB)based on the convolutional block attention module(CBAM);the discriminator adopts conventional VGG-128.Experiments show that it performs better on unpaired datasets with more noise and datasets without ground truth.