On the application of structured sparse model selection to JPEG compressed images

The representation model that considers an image as a sparse linear combination of few atoms of a predefined or learned dictionary has received considerable attention in recent years. Among the others, the Structured Sparse Model Selection (SSMS) was recently introduced. This model outperforms different state-of-the-art algorithms in a number of imaging tasks (e.g., denoising, deblurring, inpainting). Despite the high denoising performances achieved by SSMS have been demonstrated, the compression issues has been not considered during the evaluation. In this paper we study the performances of SSMS under lossy JPEG compression. Experiments have shown that the SSMS method is able to restore compressed noisy images with a significant margin, both in terms of PSNR and SSIM quality measure, even though the original framework is not tuned for the specific task of compression. Quantitative and qualitative results pointed out that SSMS is able to perform both denoising and compression artifacts reduction (e.g., deblocking), by demonstrating the promise of sparse coding methods in application where different computational engines are combined to generate a signal (e.g., Imaging Generation Pipeline of single sensor devices).