Texture-constrained Local Self-regression Super-resolution Reconstruction Algorithm

With the development of network, communication and digital technology, high-resolution images which can provide more details are becoming widespread needs of people. The purpose of image super-resolution reconstruction algorithm is to use one or more low-resolution images to obtain a clear high-resolution image. Due to the low cost and making the existing low-resolution imaging systems still used, super-resolution has always been a popular research topic.This dissertation first reviews the interpolation-based, the reconstruction-based and learning-based super-resolution reconstruction algorithms, then focuses on the self-learning super-resolution reconstruction algorithms. Based on the local self-learning super-resolution reconstruction algorithm, we research how to improve the reconstruction effects and propose a texture-constrained local self-regression super-resolution reconstruction algorithm.For the deficiency of the local self-learning super-resolution reconstruction algorithm in reconstructing the high-frequency details, firstly, we establish the mathematical model of the learning-based super-resolution reconstruction algorithm and obtain the first-order Taylor estimation of the map from the low-frequency patch to the high-frequency patch to improve the accuracy of reconstruction. Secondly, we exploit the local scaling invariance and regression to obtain the accurate expression of the first order estimation which we will use to reconstruct the high-resolution image.In searching for similar patches, to obtain the texture-similar patches for learning, we construct48filters to use textural context to express textural features and obtain the ultimate similar patches by texture-constrained search algorithm.Finally, to insure the consistency and edge-sharpness of the reconstructed high-resolution image, we exploit the corresponding prior knowledge to further optimize the reconstructed high-resolution image.Simulation experiments show that our algorithm can reconstruct reasonable high-frequency details, consequently the reconstructed images look more real and natural.