Optimum multiscale decomposition in NSCT-based single image super resolution

Abstract

Goal of image super resolution (SR) is to enhance the size of an image without upsetting the information it carries. The information around the edges may get disturbed due to image SR process. The proposed novel approach conserved the information around edges using Non-sub sample contourlet transform (NSCT)-based learning technique. The NSCT constitutes multiscale decomposition as well as multidirectional decomposition of an image. The main contribution of the paper is to optimally select multiscale decomposition level that offers minimum learning error. The smoothness of soft edges is to be potted by soft edge smoothness prior. The learning process leads to appearance of unwanted outliers which are sustained by using a robust error norm. The cost function consisting of a global constraint term and soft edge smoothness prior is optimised using Iterative Back Projection approach. The proposed method is capable to reconstruct a high-resolution image with minimum edge artefacts.

Keywords:

• Image super resolution
• Non-sub sample contourlet transform
• Soft edge smoothness prior

Acknowledgements

We would like to thank Professor Mita C. Paunwalla for fruitful discussions and kind cooperation with encouraging attitude.