Research On Image Super-resolution Reconstruction Via PCA Dictionary Learning

Image super-resolution reconstruction (SRR) technical is one of the hot research directions inthe area of digital image processing (DIP) in recent years. It can estimate one or morehigh-resolution (HR) images through one or more low-resolution (LR) ones which belong to thesame scene and have sub-pixel offsets each other, without need for improving hardware, and withthe advantage of low cost and operating easy. Therefore, this technology is widely used in satelliteimaging, medical imaging, video surveillance and other fields. In this paper, the image SRR hasbeen studied based on principal component analysis (PCA) and kernel principal component analysis(KPCA). The main work and preliminary results are as follows:Firstly, the history of SRR is reviewed. The theories basis, observation model and severalclassical SRR methods are introduced briefly.Then, analyzing the shortcoming of using over-complete dictionary in sparse model and thedeficiencies of the single-frame reconstruction, a SSR method for multiple frames via PCAdictionary learning is proposed. Firstly, the HR sample patches are divided into several classes byk-means clustering method, PCA transformation is performed on each category for dictionarylearning and the transformation matrix is made as corresponding dictionary, on which buildingsparse model. Compared with the over-complete single dictionary, these classified dictionaries areadaptive to different local image structures. Secondly, the optical flow method for motionestimation is used to search the similarity between adjacent frames, via which regularize the sparsemodel. Finally, the soft-threshold shrinkage method for solving the objective function is employedto get the target HR image. The simulation experiment to some standard test videos indicates thatthe proposed algorithm has significant improvement, compared with related algorithms.After analyzing the key issuse of KPCA and its application in DIP, a single-frame SRR methodusing KPCA is proposed to reconstruct directly the high-frequency component of the target HRimage. Firstly, based on KPCA theory, the high-frequency component of initial HR estimation ofLR image is mapped into the high dimensional feature space, and the expression of its projection inthe reproducing kernel Hilbert space is obtained. Then the corresponding pre-image in the inputspace, which is alao the high-frequency component of the target HR image, related to the projectionis sloved by means of the method based on distance constraints. Finally, the high-frequencycomponent is superimposed on the initial HR estimation to get the tatget HR image. Among them,the input space is obtained from PCA dictionary learning. That is, the PCA transformation isperformaned on each clustered high-frequency component of HR sample patches, and thetransformation matrix composed the corresponding input space of KPCA. In this way, not onlyacquiring samples belonging to a certain pattern, but also reducing the size of the sample, thus making KPCA transformation more quickly and extracting features more exactly. To futher enhancethe quality of the reconstructed image, the pre-and post-processing are performed based onobservation model. The simulation experiment to some standard test images demonstrates that theproposed algorithm has significant improvement compared with other algorithms on objective testsand subjective feeling.Finally, the main research work of this paper is summarized, the inadequacies of the study areillustrated, and the direction of futrue research is also made clear.