Optimization Of Feature Clustering And Reconstruction For Single Image Super Resolution

People always want to get higher resolution images in order to get more image information.However,using hardware imaging equipment to obtain high-resolution images has certain requirements on the imaging environment and its cost is high.Therefore,methods of using software to improve the existing image resolution have come into being.This method is collectively referred to as the image super-resolution reconstruction algorithm or technology.Theoretically,the image super-resolution reconstruction algorithm belongs to the solution of the inverse problem,and the basis of the solution mainly depends on the analytical extension theory and regularization theory.The main idea of image super-resolution algorithm based on regression learning is to obtain the model parameters by training the regression model with external high-and low-resolution training sets and then apply those parameters to the input low-resolution image to get the output high-resolution image.In this paper,in order to improve the utilization efficiency of the training set features and obtain higher reconstruction accuracy,the information of features themselves and information between features are used to improve the single-frame image super-resolution reconstruction algorithm.The main innovations are summarized as follows:Optimize the feature clustering during the training phase,including the feature reduction problem and the number of clusters K value determination problem.Based on the gradient characteristics of the image and mathematical statistical theory,a single image superresolution algorithm based on mean curvature pre-classification and Gaussian means clustering is proposed.The algorithm first uses the two-step feature of the image to solve the mean curvature and uses it as the pre-classification criterion.Then,the training set is divided into three categories to achieve the purpose of streamlining the training set.After that,the Gaussian mean clustering algorithm was used to aggregate the training set according to the categories to solve the problem of determining the K number of clusters.Finally,the corresponding regression matrix was obtained by training the ridge regression model to realize the construction of high-resolution images.The experimental results prove that the classification method based on mean curvature and the Gaussian mean clustering method proposed in this paper are feasible,and the reconstructed image quality by using the proposed algorithm is better than that of using other mainstream algorithms.Optimize the feature reconstruction at the reconstruction stage,including the feature metric selection problem and the regularization selection problem: Based on the inverse problem and the similarity theory between features,a single image super-resolution based on the Mahalanobis metric and weighted regularization is proposed.The algorithm first introduces the Mahalanobis metric formula to measure the feature space,and uses the Lipschitz function metric learning method to learn the Mahalanobis metric formula that corresponds to each feature set.These methods solve the problem that the training feature set cannot be well measured when Euclidean distance is used in the feature measurement.Then,the Lp regularization is used in the regression model,and the adaptive similarity coefficient is introduced into the regularization,so that the mapping matrix can be obtained by using the prior knowledge of feature space similarity in the training set.Finally,in the reconstruction stage,according to the Mahalanobis metric corresponding to the training features,the optimal mapping matrix is matched for the input low-resolution features,and then a highresolution image is reconstructed.The experimental results show that the method based on Mahalanobis metric for feature matching and the use of weighted regular terms in the model are effective.At the same time,the results prove that the proposed algorithm in this paper obtains better reconstruction quality.