Automatic Patch Classification and Diffusion in Adaptive Super-Resolution Based on Optimal Recovery

Luis Alonso Ponce Loya, University of Texas at El Paso


Multi-Frame image restoration is a form of Super-Resolution (SR) which consists of combining multiple Low-Resolution (LR) images in order to reconstruct a single High Resolution (HR) image. In this thesis we improved a spatially-adaptive SR scheme, based on the framework of Optimal Recovery (OR), whereby the block-by-block processing is done based on the properties of the corresponding local LR image data patches. The bandwidth parameter is adapted based on the local variance of each corresponding block from all LR images that contribute to the HR image block to be reconstructed. Also, an optimal regularization parameter needs to be chosen for the reconstruction of each HR output block. Some improvements achieved over a previous MATLAB implementation include: remove anti-aliasing filter to generate LR images so we can recover HR frequencies, implement a different shift registration for LR images, improve the selection of parameters and block classes according to the input image, bypass the use of Generalized Cross-Validation (GCV) for computation efficiency, implement a variable block size selection, allow different input image sizes, apply code debugging techniques to make the code more efficient in time, and solve image boundaries problems. Furthermore, we extended this approach to obtain estimated gradients as well as to generate edge maps by using OR theory. We also integrated an anisotropic diffusion method in two different ways, as a global post processing step on the whole reconstructed HR image or as a local reconstructed patch post processing step. Final results show significant improvement over previous versions of this approach and over many other SR algorithms of low computational complexity. ^

Subject Area

Electrical engineering

Recommended Citation

Ponce Loya, Luis Alonso, "Automatic Patch Classification and Diffusion in Adaptive Super-Resolution Based on Optimal Recovery" (2017). ETD Collection for University of Texas, El Paso. AAI10285312.