Reproduce “Fidelity-preserving Learning-Based Image Compression: Loss Function and Subjective Evaluation Methodology” on OpenResearch

About this paper

Shima Mohammadi, Yaojun Wu, João Ascenso

This work introduces a learning-based image compression method featuring a novel loss function that integrates perceptual and adversarial components to enhance visual quality while preserving image fidelity. A specialized triplet subjective evaluation methodology is also presented, demonstrating that the perceptually optimized codec offers improved perceived fidelity, particularly at lower bitrates, compared to conventionally optimized approaches.

Key insights

• It is possible to achieve perceptually enhanced image compression while preserving fidelity by carefully balancing adversarial and perceptual losses with traditional distortion metrics during network training.
• Standard subjective evaluation methods are often insufficient for accurately assessing fidelity in perceptually optimized images; a direct comparison against a reference in a triplet setup can effectively capture subtle fidelity differences.
• The benefits of perceptual optimization for fidelity are most pronounced at lower bitrates, where compression artifacts are more apparent, with diminishing returns at higher bitrates.

Summary by alphaXiv.