Adaptive sparse modeling in spectral & spatial domain for compressed image restoration

Abstract

Block discrete cosine transform (BDCT) is an indispensable component of modern image and video coding standards, specifically for its decorrelation and superior energy compaction aspects. BDCT typically employs block-specific quantization, which results in unpleasant compression-blocking artifacts which predominates at low bit rates. The proposed method aims to minimize these blocking artifacts to generate high-quality images under the framework of the alternating direction method of multipliers (ADMM) optimization. The proposed method exploits the local structures identified and extracted via wavelet-patch-based sparse representation and non-local self-similarity identified and extracted via group-based sparse representation, which are subsequently combined optimally employing ADMM. Moreover, the method uses a Gaussian quantization noise model, which allows a more precise and reliable assessment. An adaptive regularization parameter is used, which integrates spectral and spatial domain sparse representations with multi-resolution dictionaries and PCA-based dictionary. The proposed algorithm improves the overall practicality of the process and outperforms existing methods in terms of objective measures like structural similarity index measure, peak signal-to-noise ratio and visual perception.