Edge preserving super-resolution with details based on similar texture synthesis

In this paper, we propose a method for single image super-resolution that integrates reconstruction-based and example-based super-resolution. In image enlargement, jagged and blurred artifacts are often generated, particularly along the edges. Under large magnification, the region containing many edges of fine texture becomes blurred, and the reconstructed image looks unnatural. The proposed method consists of (i) a reconstruction process with local regularization and non-local regularization and (ii) a detail estimation process based on texture synthesis. Local regularization suppresses artifacts by estimating a pixel of interest in a local area. Non-local regularization reduces the noise by estimating a pixel of interest as a weighted average of similar pixels in a non-local area. Texture synthesis recovers fine texture by incorporating the optimal patch selected from a dictionary. For example-based super-resolution, patch size based on an importance map is used. These processing steps are performed to preserve the original edge patterns. In addition, back projection is applied to obtain the optimal image. Finally, we show some simulation results with less jagged and blurred artifacts compared to those of conventional methods.