Research map / Reinforcement learning / Visual and representation RL
Fully Convolutional Network with Multi-Step Reinforcement Learning for Image Processing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add convGRU to the policy network (ablation) | pixelRL denoising · BSD68 Gaussian denoising | PSNR (dB): 31.17 -> 31.26 (σ=15); 28.75 -> 28.83 (σ=25); 25.78 -> 25.87 (σ=50) | Helped |
| Add reward map convolution (RMC) (ablation) | pixelRL denoising · BSD68 Gaussian denoising | PSNR (dB): 31.26 -> 31.40 (σ=15); 28.83 -> 28.85 (σ=25); 25.87 -> 25.88 (σ=50) | Mixed |
| Apply test-time augmentation (ablation) | pixelRL denoising with convGRU and RMC · BSD68 Gaussian denoising | PSNR (dB): 31.40 -> 31.49; 28.85 -> 28.94; 25.88 -> 25.95 | Helped |
| Add reward map convolution (RMC) (ablation) | pixelRL denoising · BSD68 salt-and-pepper denoising | PSNR (dB): 37.86 -> 38.46; 29.26 -> 29.78; 23.54 -> 23.78 | Helped |
| Apply test-time augmentation (ablation) | pixelRL denoising with convGRU and RMC · BSD68 salt-and-pepper denoising | PSNR (dB): 38.46 -> 38.82; 29.78 -> 29.92; 23.78 -> 23.81 | Helped |
| Add reward map convolution (RMC) (ablation) | pixelRL image restoration · Image restoration test set | PSNR (dB): 29.50 -> 29.97 | Helped |
| Add reward map convolution (RMC) (ablation) | pixelRL local color enhancement · Local color enhancement styles | mean L2 error: 6.75 -> 6.69 (Foreground Pop-Out); 6.17 -> 5.67 (Local Xpro); 6.44 -> 6.41 (Watercolor) | Helped |
| Add bilateral filter with σc=1.0 to the basic filter action set (ablation) | pixelRL denoising · BSD68 Gaussian denoising | PSNR (dB): 29.82 -> 29.93 (σ=15); 27.60 -> 27.81 (σ=25); 25.20 -> 25.30 (σ=50) | Helped |
| Add pixel value +=1 and -=1 actions (ablation) | pixelRL denoising with bilateral filter · BSD68 Gaussian denoising | PSNR (dB): 29.93 -> 30.72 (σ=15); 27.81 -> 28.25 (σ=25); 25.30 -> 25.59 (σ=50) | Helped |
| Add bilateral filter σc=0.1 and Gaussian filter σ=0.5 (ablation) | pixelRL denoising action set · BSD68 Gaussian denoising | PSNR (dB): 30.72 -> 31.26 (σ=15); 28.25 -> 28.83 (σ=25); 25.59 -> 25.87 (σ=50) | Helped |
| Add advanced filters such as guided filter or non-local means (ablation) | pixelRL denoising action set · Image denoising | PSNR | No clear effect |
| Use pixelRL rather than CNN | BSD68 denoising · Salt-and-pepper density=0.5 and 0.9 | PSNR (dB): 29.40 -> 29.92 (density=0.5); 23.76 -> 23.81 (density=0.9) | Helped |
Closest papers by meaning
- RL-SeqISP: Reinforcement Learning-Based Sequential Optimization for Image Signal Processing (AAAI 2024 · similarity 0.87)
- Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.84)
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (NeurIPS 2026 · similarity 0.83)
- Learning to Control Camera Exposure via Reinforcement Learning (CVPR 2024 · similarity 0.83)
- Goal Conditioned Reinforcement Learning for Photo Finishing Tuning (NeurIPS 2024 · 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.