Research On Learning-based Image Super-resolution Reconstruction Algorithms

In the process of capturing images, by the limit of the imaging condition andway, imaging system cannot get all the information in the original scene. In thepremise of super resolution reconstruction technology without changing thehardware conditions, the economic and effective way to solve the problem ofimaging is using the method of image processing, reconstructing systeminformation out of the cut-off frequency to obtain the image whose information ishigher than the resolution of imaging systems, The technology can improve thespatial resolution of the image without hardware involvement and cost low, Inmany areas such as remote sensing, medical imaging, video surveillance, digitalmedia has been initially applied, with a huge potential for development.In this paper, combined with the excellent characteristics of sparse theory,studies the Learning-based Image Super-Resolution Reconstruction algorithm,including the following:The feature extraction and dimension reduction in the over-completedictionary training phase is improved, In the feature extraction using thecombination of the two order derivative and the gradient process, construct a newdirection of descent, with a descent direction of new design an algorithm,improved the gradient method. The new algorithm is better than the convergencerate of the conjugate gradient method is fast, better results in feature extraction.The improvement Two-dimensional Principal Component Analysis (2DPCA)algorithm is applied on these vectors, can also eliminate the correlation imagerows and columns, experiments show that the method of image reconstructioneffect is better, faster.Research the image super resolution reconstruction algorithm based onadaptive sparse domain selection. Firstly, introduce the image super resolutionreconstruction and adaptive sparse domain selection. Then, Research the Super-resolution by Adaptive Sparse Domain Selection algorithm, get the targetmodel of super-resolution reconstruction., add two adaptive regularizationconstraint items into this algorithm, they are Adaptive selectionautoregressive(AR) model regularization and non-local self-similarityregularization. The combination of the two kinds of models add to the adaptivesparse domain selection model to get the SR reconstruction algorithm, Make fulluse of prior knowledge of image, and establishes the model of SR reconstructionof relatively complete.