Revisit of JPEG Decompression with Artifact Suppression by Total Variation and its Variants
Yanhao Li, Quentin Bammey, Marina Gardella, Rafael Grompone von Gioi, Miguel Colom, Jean-Michel Morel
⚠ This is a preprint. It may change before it is accepted for publication.

Abstract

The question of JPEG artifacts reduction has existed since the birth of JPEG compression, and gained importance due to the popularity of JPEG compression. Suppression approaches are called idempotent when they guarantee the preservation of quantized DCT coefficients both before and after recompression using the same quantization matrix. We briefly revisit the idempotent JPEG decompression method based on total variation (TV) minimization using three variant costs: classical TV, adapted TV and Dirichlet integral. We review the constrained optimization problem of the JPEG decompression and its solution by gradient descent with projection onto a convex set, followed by a detailed description of its implementation. Both quantitative and qualitative experiments demonstrate the efficacy of this decompression method in eliminating artifacts and the over-optimization issue that smooths out original textures. Additionally, we investigate the performance differences among the three variant cost functions.

Download