Research On Image Super-resolution Method Based On Mapping Kernel Learning

With the development of computer technology and artificial intelligence technology,people have more capable to analyze and utilize 2D signals such as images and videos.For example,police officers can use the tracking algorithm to process surveillance video to monitor and track suspicious targets,medical staffs can use detection algorithm to check a large number of CT maps and so on.However,due to the constraints and influences of imaging equipment and transmission noise,the quality of the acquired images often fails to meet the requirements of the tasks.Therefore,in the face of complicated small-sized and low-quality image data,there is an urgent need for accurate and efficient high-definition image restoration.Image Super-resolution(SR)technology uses algorithms to break through the limitations of imaging and transmission processes,and reconstructs low-resolution(LR)images into clear high-resolution(HR)images.SR technology is a classic problem in the field of digital image processing.The development of SR technology is of great significance for multimedia applications,medical image processing,public security and other fields.Recently,with the development of machine learning,SR methods based on local-patch mapping-kernel learning have become a hot research direction.These methods usually include three key steps:external sample classification,class mapping-kernel learning,and image post-processing.This paper improves the above steps and propose some novel high reconstruction quality and high-speed SR methods and the main innovative research results are as follows:(1)An image SR algorithm combining sample classification scheme based on Sample Individual Mapping Kernel(SIMK)and Ill-Sample Remove Mechanism is proposed.The algorithm improves the classification rules of LR-HR patch pairs based on LR patch gradient features in the current algorithm.This method classifies external samples by using SIMK and learning the naive Bayesian classifier to classify LR patches,thereby improving the reconstruction accuracy.In the mapping learning process,the algorithm also uses the Ill-Sample Remove Mechanism to remove the samples away from the center of the classified sample set,thereby improving the validity of the learned mapping kernels.Comparative experiments show that the proposed method achieves an improvement in both visual and performance compared to the current state-of-the-art linear mapping-based SR methods;(2)A real-time and efficient SR algorithm combining sample classification decision tree based on individual sample mapping vector and mapping kernel learning stage based on nonlinear regression is proposed.The algorithm can both guarantee the execution speed and reconstruction quality at patch representation and mapping learning stages.In the training phase,the algorithm selects the mapping vector of each sample as the basis,and uses the fractional norm clustering method to achieve more accurate external sample classification,and constructs a binary classification decision tree to complete the sample classification layer by layer.The samples of each parent node are divided into two child nodes based on the map vectors.In order to ensure the validity of the node branch,this algorithm designs an accurate leaf node marking method to ensure the validity of the node branch.At the leaf node,the algorithm learns a lightweight full-connected network to achieve pixel-level HR block reconstruction,which ensures both speed and reconstruction quality.Comparative experiments show that the proposed method has stronger speed and visual advantages than the current advanced neural network-based super-resolution methods.(3)A post-processing algorithm for SR image texture optimization based on external example learning and spatial-frequency feature-based classification is proposed.The algorithm collects LR-HR texture patch pairs from the training set and learns corresponding optimized mapping kernels for different types of texture patches.The algorithm performs feature analysis on LR texture patches based on the features of space and frequency domains.In order to ensure that there are no noise samples in the collected texture tile pairs,this algorithm trains an SVM-based selector,which uses the gradient features of LR texture tiles and fractional Fourier spectral features as support vectors.As a post-processing algorithm,the algorithm can further improve the existing SR algorithms in terms of visual and performance.This paper also designs a mapping kernel learning based super-resolution system,which based on the proposed methods.After testing,the system basically satisfies the user's need for real-time and multiple magnification super-resolution of natural images.