Research map / Image generation / Image restoration
Enhancing Quality of Compressed Images by Mitigating Enhancement Bias Towards Compression Domain
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies an enhancement bias toward the compression domain in GAN-based compressed-image quality enhancement and mitigates it with a compressed-image-conditioned discriminator plus a VGG-feature domain-divergence regularization.
Problem
GAN-based quality enhancement of compressed images makes outputs resemble compressed inputs, because discriminators treat compression artifacts as realistic, degrading perceptual quality.
Approach
Use a conditional discriminator conditioned on the compressed image (exploiting conditional independence of raw and enhanced images given the compressed input) and a domain-divergence loss that pushes the enhanced domain away from the compressed domain toward the raw domain; model-agnostic and adds no inference cost.
Key results
- Ours with Real-ESRGAN [44] reduces FID from 10.6 to 8.77 on BPG QP37 DIV2K and raises PSNR from 29.8 to 30.6 dB (Table 3 variants).
- Ours with ESRGAN [43] lowers FID to 7.50 and improves PSNR to 29.6 dB, with BD-BR of -24.3% FID and -21.8%+ PSNR gains.
- Bias analysis shows baselines have enhancement-bias deviations of -45% to -98%, which become positive (+11% to +19%) after the method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add conditional discriminator and domain-divergence regularization | [44] · DIV2K BPG QP 37 | FID: 10.6 -> 8.77 | Helped |
| add conditional discriminator and domain-divergence regularization | [44] · DIV2K BPG QP 37 | PSNR (dB): 29.8 -> 30.6 | Helped |
| add conditional discriminator only (ablation) | [43] · DIV2K BPG QP 37 | FID: 8.84 -> 8.05 | Helped |
| add conditional discriminator only (ablation) | [43] · DIV2K BPG QP 37 | PSNR (dB): 29.0 -> 29.4 | Helped |
| add domain-divergence regularization only (ablation) | [43] · DIV2K BPG QP 37 | FID: 8.84 -> 7.67 | Helped |
| add domain-divergence regularization only (ablation) | [43] · DIV2K BPG QP 37 | PSNR (dB): 29.0 -> 29.2 | Helped |
| add both conditional discriminator and domain-divergence regularization (ablation) | [43] · DIV2K BPG QP 37 | FID: 8.84 -> 7.50 | Helped |
| add both conditional discriminator and domain-divergence regularization (ablation) | [43] · DIV2K BPG QP 37 | PSNR (dB): 29.0 -> 29.6 | Helped |
| add conditional discriminator only (ablation) | [44] · DIV2K BPG QP 37 | FID: 10.6 -> 9.21 | Helped |
| add conditional discriminator only (ablation) | [44] · DIV2K BPG QP 37 | PSNR (dB): 29.8 -> 30.0 | Helped |
| add domain-divergence regularization only (ablation) | [44] · DIV2K BPG QP 37 | FID: 10.6 -> 9.70 | Helped |
| add domain-divergence regularization only (ablation) | [44] · DIV2K BPG QP 37 | PSNR (dB): 29.8 -> 30.4 | Helped |
Surprising findings
- Existing discriminators judge compressed images as more realistic than raw images (average realism score for compressed exceeds raw).
- Real-ESRGAN enhanced outputs have FID (10.6) higher than the compressed input (9.73), indicating post-enhancement fidelity degradation.
Limitations
- Evaluated only on BPG and JPEG compression and the DIV2K/Flickr2K data.
- Depends on GAN-based SR baselines; gains reported relative to re-trained baselines with a single fixed hyperparameter setting.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: DIV2K validation set Flickr2K
Metrics: FID LPIPS PSNR AHIQ CLIP-IQA DISTS HyperIQA MUSIQ NIQE PI
Baselines: AR-CNN DCAD DnCNN CBDNet RDN RBQE MPRNet ESRGAN Real-ESRGAN
Builds on: Real-ESRGAN ESRGAN conditional GAN SRGAN
Closest papers by meaning
- Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment (CVPR 2024 · similarity 0.84)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.84)
- Make Lossy Compression Meaningful for Low-Light Images (AAAI 2024 · similarity 0.84)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.83)
- Correcting Diffusion-Based Perceptual Image Compression with Privileged End-to-End Decoder (ICML 2024 · similarity 0.83)
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.