Enhancing Quality of Compressed Images by Mitigating Enhancement Bias Towards Compression Domain

TL;DR

This method employs a conditional discriminator with the compressed image as a key condition, and then incorporates a domain-divergence regularization to actively distance the enhancement domain from the compression domain, and brings the enhancement domain closer to the raw domain.

Abstract

Existing quality enhancement methods for compressed images focus on aligning the enhancement domain with the raw domain to yield realistic images. However, these methods exhibit a pervasive enhancement bias towards the compression domain, inadvertently regarding it as more realistic than the raw domain. This bias makes enhanced images closely resemble their compressed counterparts, thus degrading their perceptual quality. In this paper, we propose a simple yet effective method to mitigate this bias and enhance the quality of compressed images. Our method employs a conditional discriminator with the compressed image as a key condition, and then incorporates a domain-divergence regularization to actively distance the enhancement domain from the compression domain. Through this dual strategy, our method enables the discrimination against the compression domain, and brings the enhancement domain closer to the raw domain. Comprehensive quality evaluations confirm the superiority of our method over other state-of-the-art methods without incurring inference overheads.

Figures (10)

• Figure 1: Top: Frechet inception distance (FID) heusel_gans_2017 scores between enhancement, compression, and raw domains on the DIV2K validation set agustsson_ntire_2017. Bottom: Visualization of residual to the compressed image. The results illustrate the enhancement bias towards the compression domain. Our method effectively mitigates this bias, bringing the enhancement domain closer to the raw domain.
• Figure 2: Visual demonstrations of realism scores for images as evaluated by wang_real-esrgan_2021. The results highlight that existing methods overlook significant compression artifacts in realism evaluation.
• Figure 3: Similarity scores between compression, raw, and enhancement domains. Lower FID and LPIPS scores indicate greater similarity. The horizontal deviation of each vertex relative to the centroid of its base, calculated with float precision on original data, is also shown. The results underscore the enhancement bias, demonstrating a closer alignment of the enhancement domain with the compression domain than with the raw domain.
• Figure 4: Residual comparisons between enhanced images wang_real-esrgan_2021 and their compressed/raw counterparts. These visualizations reveal a stronger resemblance of the enhanced images to their compressed counterparts, as indicated by the weaker residual.
• Figure 5: Causal relationships among the raw, compressed, and enhanced images, where $f_c$ and $f_e$ denote the image compression and quality enhancement respectively.