Research On Image Super-Resolution Reconstruction Algorithm Based On Feature Classification And Sparse Representation

Due to the quality of the imaging sensor or the interference of the external factors,people perhaps get images with low resolution in reality,which cannot meet the needs of people's daily life or production environment.Image super-resolution is a corresponding solution method,which aims to use some algorithms to improve the resolution of the image,and as much as possible to recover image details for keeping the image clear.Currently,image super-resolution is applied widely in many fields,such as medical imaging,intelligent surveillance,remote sensing and so on.In recent years,as the development of computer technology,super-resolution has also made great breakthroughs,especially the learning-based super-resolution.The learning-based method can easily generate high-frequency details of image that are not available in the input image,which is why so many researchers pay attention to and study it.We mainly focus on the methods of dictionary learning for super-resolution.The main contents of research in this thesis are as follows:1.The background and significance of image super-resolution are described,and then the current research status of this technology is introduced.In addition,sparse representation and feature classification are emphasized,which are the foundation of super-resolution in our works.In the stage of feature classification,some image-related features are extracted,and then they are used to construct a classification decision tree,which will be used to classify the image patches.The experiment shows it works well.2.The traditional methods of dictionary learning directly use the entire training set to train the dictionaries.Therefore,the atoms with different kinds of features are arranged indifferently in the dictionary,which will not be good for the representation and reconstruction of image patches.So a method of super-resolution based on feature classification and dictionary mapping is proposed in this thesis.Firstly,this method uses the classification decision tree to classify the image patches,then it uses the low-resolution training set in each category to train the low-resolution dictionary.Next,the high-resolution dictionary is obtained by using the method of dictionary mapping based on the invariance of the high-and low-resolution coefficients.The high-resolution dictionary will be used in the stage of image reconstruction.3.In fact,the high-resolution patch's coefficient over the high-resolution dictionary is not strictly same to the low-resolution patch's coefficient.Therefore,a method of super-resolution based on independent dictionary training and the optimization of MAP framework is proposed in this thesis.First of all,this method separately trains the high-and low-resolution dictionaries,and get the corresponding coefficient matrices.For a given low-resolution patch to be constructed,this method learns a coefficient mapping using the current high-and low-resolution coefficient matrix,in order to obtain its potential real high-resolution coefficients.Similarly,the coefficients are used in the stage of image reconstruction.Next,the non-local self-similarity and the auto-regression regularization are added into the MAP framework to optimize the image,which works well.