Research On Image Super-resolution Reconstruction Based On Deep Learning

Image super-resolution reconstruction is an image processing technology which processes and analyzes a low-resolution image or a series of images by a computer to recover a desired high-resolution image,it has become one of the hot topics in the field of machine learning,pattern recognition,and other related research fields as people's demand for image accuracy increases.Currently,super-resolution reconstruction algorithms are divided into two categories based on interpolation and learning algorithms.Interpolation-based algorithms are simple and quick,but they cannot meet people's growing needs of image quality.The vigorous development of deep learning has enabled researchers to apply deep learning algorithms to image super-resolution reconstruction and achieved better reconstruction results than interpolation algorithms.This dissertation is based on the deep learning algorithm to research single image super resolution.This dissertation focuses on deep convolutional networks,studying traditional interpolation reconstruction,sparse coding reconstruction algorithms,researching their basic ideas and deficiencies,taking their essence,discarding the dross,and proposing improved algorithms.The reconstruction performance and robustness have been improved.The main contributions are summarized as follows:(1)The traditional sparse-coded SR algorithm has the advantages of being simple and fast,but it is easy to produce the impurity and blur phenomenon after restoring the image result.For the shortcomings of sparse-coded SR algorithm,such as poor recovery and many missing images.Combined with the strong learning ability of deep network,it is introduced on the basis of traditional sparse coding algorithm,and a better image SR method is developed while better results are obtained on Set5 and other data sets.The experimental results on the data set show that the improved method in this dissertation can effectively improve the image reconstruction quality.(2)In theory,the deeper the depth network used by the image SR algorithm is,the better the quality of the reconstructed HR image.Based on the VDCN network structure,a new high-precision single-image super-resolution method is proposed by replacing the activation function ReLU with PReLU and deepening the number of network layers for the disadvantages of poor SR effect on large-scale factors of VDCN,which called Super Deep Convolutional Network.And we do experiments on data sets such as Set5.Experimental results show that the proposed method in this dissertation outperforms the VDCN method in large-scale factors.(3)Most existing SR methods based on deep learning use only a single model to generate HR results,which can easily lead to local optimization and poor image processing.In view of the above-mentioned disadvantages of the existing deep learning SR algorithms,based on the homogeneous network of ensemble ideas,the ensemble of heterogeneous network was proposed,and good experimental results are obtained on Set5 and Set14 data sets.