A nonparametric Bayesian method of structural saliency dictionary learning for image compression.

Juan Wang,Xiaoming Tao,Chunxiao Jiang,Shaoyang Li,Jianhua Lu
DOI: https://doi.org/10.1109/globalsip.2017.8309151
2017-01-01
Abstract:Region-of-interest (ROI) coding is considered to be effective in preserving high quality of regions which are rather significant for an image with low bit-rate. Existing methods suffer from two problems that potentially limit their practical implementations: firstly, manual ROIs annotation is time consuming and sometimes impractical; secondly, predefined coding transform basis is not optimal due to the distinct structural patterns between ROIs and non-ROIs. In this paper, we address these limitations by: 1) automatically extracting ROIs by the saliency detection technique; 2) employing a nonparametric Bayesian method to learn the distinct structural patterns. Furthermore, we design a Joint-OMP sparse representation algorithm to compress the images based on the learned dictionaries. Experimental results demonstrate that our method outperforms the existing ROI-based image compression methods in terms of PSNR, while exhibiting improvements on visual quality of the reconstructed images.
What problem does this paper attempt to address?