Research Paper#Image Super-Resolution, Reinforcement Learning, Reward Models🔬 ResearchAnalyzed: Jan 3, 2026 16:23
FinPercep-RM: Fine-grained Reward Model for Real-world Super-Resolution
Analysis
This paper addresses the limitations of traditional Image Quality Assessment (IQA) models in Reinforcement Learning for Image Super-Resolution (ISR). By introducing a Fine-grained Perceptual Reward Model (FinPercep-RM) and a Co-evolutionary Curriculum Learning (CCL) mechanism, the authors aim to improve perceptual quality and training stability, mitigating reward hacking. The use of a new dataset (FGR-30k) for training the reward model is also a key contribution.
Key Takeaways
- •Proposes FinPercep-RM to address the insensitivity of traditional IQA models to local distortions.
- •Introduces the FGR-30k dataset for training the FinPercep-RM.
- •Employs a Co-evolutionary Curriculum Learning (CCL) mechanism to stabilize training.
- •Focuses on improving perceptual quality and mitigating reward hacking in RL-based ISR.
Reference
“The FinPercep-RM model provides a global quality score and a Perceptual Degradation Map that spatially localizes and quantifies local defects.”