The Research On Image Super-resolution Method Based On End-to-End Deep Learning

In this era of intelligence and informationization,images play a vital role in the process of information transmission,and people have increasingly higher requirements for image quality.However,image acquisition and imaging are often affected by factors such as imaging equipment,motion blur,and noise,resulting in lower resolution of the acquired images and failing to meet the growing demand for high-quality images.Therefore,converting lowresolution images to high-resolution images with more pixel density and detail is crucial,which is one of the important courses in the field of image processing.Super-resolution reconstruction technology is to use a series of algorithms to process low-resolution images to reconstruct high-resolution images without changing the hardware equipment.The traditional image super-resolution algorithm has insufficient prior information and the reconstruction effect is not ideal.The introduction of deep learning methods can learn richer image information,which greatly improves the super-resolution reconstruction effect.Based on the research of existing deep learning methods in the field of superresolution,this paper improves on it.The main research work is as follows:(1)Deep learning methods can effectively improve the effect of image super-resolution reconstruction.At present,deep learning-based image super-resolution reconstruction methods have problems such as slow network convergence,poor image texture reconstruction,and inability to perform multi-scale reconstruction.This paper proposes a deep convolutional neural network based on residual learning.The network consists of two subnets,a local residual network and a global residual network.The joint training of the two subnets increases the network width and restores the different features of the image.Among them,the local residual network is divided into three modules of feature extraction,upsampling and multi-scale reconstruction,and the global residual network is divided into feature extraction and image reconstruction modules.In the network,the convolutional layers are densely connected through residual dense blocks to fully extract local features at various levels,and multi-scale convolutional layers are used to obtain rich image context information.Multi-scale image reconstruction is implemented using progressive upsampling,and residual learning is introduced to speed up the convergence speed during training.The method in this paper has been greatly improved in terms of subjective vision and objective quantification,which can effectively improve the quality of image reconstruction and better preserve edge details in the image.(2)Aiming at the shortcomings of the existing SR methods,such as the poor performance of large-scale factor reconstruction and the incomplete rendering of texture details,this paper proposes an SR method based on fractal coding and depth residuals.Fractal coding technology is an effective tool for describing image texture.First,quadtree fractal coding is performed on the original image,and the similarity relationship between the range block and the domain block is constructed,and an iterative function system of the image is found.Then,through this iterative function,super-resolution fractal decoding and reconstruction are performed on the attractor to obtain a preliminary fractal interpolation image.Finally,a deep residual network is constructed to estimate the enlarged fractal coded collage error,and it is accumulated as an error compensation term in the interpolated image to obtain the final reconstructed image.The network structure is obtained by joint training of deep and shallow networks,and residual learning is introduced to greatly improve the convergence speed and reconstruction accuracy of the network.Compared with the prior art methods,the method in this paper shows good performance in both quantitative and qualitative aspects,can highlight subtle edges and vivid textures,and also presents large scale factors better.