Research map / Image generation / Image restoration
Distort-and-Recover: Color Enhancement Using Deep Reinforcement Learning
Jongchan Park, Joon-Young Lee, Donggeun Yoo, In So Kweon
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use distort-and-recover DRL instead of Pix2Pix on distorted-reference training pairs | Color enhancement on MIT-Adobe FiveK Random 250 · Random 250 | mean L² error: 14.46 -> 12.15 | Improved |
| Use distort-and-recover DRL instead of Pix2Pix on distorted-reference training pairs | Color enhancement on MIT-Adobe FiveK Random 250 · Random 250 | SSIM: 0.699 -> 0.910 | Improved |
| Add contextual VGG features to the Lab histogram (ablation) | Color enhancement agent · Random 250 | mean L² error: 12.30 -> 10.99 | Improved |
| Use a CIELab histogram with context instead of six color features (ablation) | Color enhancement agent · Random 250 | mean L² error: 12.53 -> 10.99 | Improved |
| Replace the CIELab histogram with RGB-L histogram (ablation) | Color enhancement agent · Random 250 | mean L² error: 10.99 -> 12.62 | Hurt |
| Replace the CIELab histogram with TinyImage (ablation) | Color enhancement agent · Random 250 | mean L² error: 10.99 -> 15.07 | Hurt |
| Augment Pix2Pix training pairs 10 times instead of using no augmentation (ablation) | Pix2Pix with distort-and-recover training · Random 250 | mean L² error: 14.46 -> 13.79 | Improved |
| Increase Pix2Pix augmentation from 10 times to 20 times (ablation) | Pix2Pix with distort-and-recover training · Random 250 | mean L² error: 13.79 -> 13.59 | Improved |
| Increase Pix2Pix augmentation from 20 times to 30 times (ablation) | Pix2Pix with distort-and-recover training · Random 250 | mean L² error: 13.59 -> 13.60 | No clear effect |
| Train with Nashville-filtered reference images using distort-and-recover | Color enhancement agent · MIT-Adobe FiveK Random 250 (Nashville style) | mean L² error: 27.78 -> 17.87 | Improved |
| Use the proposed method instead of Pix2Pix trained with Shutterstock references | Color enhancement user study · MIT-Adobe FiveK user study | Mean score: 2.09 -> 2.86 | Improved |
| Use the proposed method instead of Lightroom auto adjustments | Color enhancement user study · MIT-Adobe FiveK user study | Mean score: 2.74 -> 2.86 | Improved |
Closest papers by meaning
- Crafting a Toolchain for Image Restoration by Deep Reinforcement Learning (CVPR 2018 · similarity 0.88)
- Fully Convolutional Network with Multi-Step Reinforcement Learning for Image Processing (AAAI 2019 · similarity 0.88)
- RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.86)
- Goal Conditioned Reinforcement Learning for Photo Finishing Tuning (NeurIPS 2024 · similarity 0.86)
- Learning to Control Camera Exposure via Reinforcement Learning (CVPR 2024 · similarity 0.85)
- Close Imitation of Expert Retouching for Black-and-White Photography (CVPR 2024 · similarity 0.84)
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.
