RGSUD: Reward-Guided Self-Reinforcement for Unpaired Image Deraining

Forum topic · 小凯 · 2026-05-04

Summary

This post introduces RGSUD (Reward-Guided Self-Reinforcement Unsupervised Deraining), a framework from the paper "Unpaired Image Deraining Using Reward-Guided Self-Reinforcement Strategy" (arXiv:2605.00719). Unsupervised image deraining struggles because real paired rainy/clean images are hard to obtain and synthetic rain is unrealistic, leaving models without strong constraints during training. RGSUD addresses this by exploiting high-quality deraining results that occasionally emerge during training. It works in two stages: reward recycling, where a reward function identifies good outputs based on rain-free image statistics (e.g., gradient distributions), semantic consistency, and naturalness; and self-reinforcement, where these identified high-quality outputs serve as pseudo-targets to guide further optimization, creating a positive feedback loop. The approach resembles intrinsic reward in reinforcement learning and requires no external supervision. The post explains why rain's variability breaks purely statistical unsupervised constraints, discusses the philosophy of capturing emergent good outcomes during training, and offers practical questions practitioners can ask when applying reward-guided self-reinforcement to other unsupervised or self-supervised learning tasks.

> Paper: Unpaired Image Deraining Using Reward-Guided Self-Reinforcement Strategy > Authors: Yinghao Chen, Yeying Jin, Xiang Chen, Yanyan Wei, Ziyang Yan, Yaowen Fu > arXiv: 2605.00719 | 2026-04-30

1. The "Rain Keeps Falling" AI Dilemma

Imagine an AI system that removes rain streaks from photos.

The traditional approach:

Collect paired "rainy/clean" images
Supervised learning: rainy input → clean output

But in reality:

True paired rainy/clean images are hard to obtain
Synthetic rain is unrealistic
Real-world rain varies endlessly

Unsupervised deraining sounds ideal: no paired data needed. But without strong constraints during training, models struggle to converge.

2. The Core Challenge of Unsupervised Deraining

The complexity of rain:

Droplets of different sizes
Varying streak density
Different directions
Mixing with background textures
Dynamic scenes are even harder

The unsupervised learning dilemma:

No "correct answer" for guidance
The network doesn't know what a "good" derained result looks like
Prone to local optima
Produces blurry or over-smoothed results

Key insight: even without paired supervision, high-quality deraining results occasionally emerge during training.

The question is: how to recognize and exploit these "lucky accidents"?

3. RGSUD: Reward-Guided Self-Reinforcement

The paper proposes RGSUD (Reward-Guided Self-Reinforcement Unsupervised Deraining):

Core idea: > Treat high-quality results that occasionally appear during training as "rewards," and recycle these rewards to guide subsequent optimization.

Two-stage strategy:

Stage 1: Reward Recycling

During training, the network occasionally produces good deraining results
A reward function is designed to identify these good results
The reward is based on:
Statistical properties of rain-free images (e.g., gradient distributions)
Semantic consistency (main objects unchanged after deraining)
Naturalness (the result looks like a real rain-free image)

Stage 2: Self-Reinforcement

The identified high-quality results serve as "pseudo-targets"
The network learns to generate more similar high-quality results
A positive feedback loop: good results → more good results

It's like a student solving problems without an answer key, but occasionally discovering a particularly elegant solution. Record the elegant solutions, analyze why they work, then deliberately repeat that "goodness."

4. Why Does "Self-Reward" Work?

Traditional unsupervised methods assume:

The data distribution contains sufficient information
The network can learn from statistical regularities alone

But rain's complexity breaks this assumption:

Rain distribution is coupled with the background
Pure statistical constraints are not enough

RGSUD's innovations:

Leveraging training dynamics: rather than statically learning the data distribution, dynamically exploit "emergent" quality during training
Self-discovery: the network discovers "what is good" on its own
No external supervision needed: the reward function is based on intrinsic image properties

This resembles "intrinsic reward" in reinforcement learning — not given externally, but generated by the system itself.

5. Finding Necessity in Chance

> "The pleasure of scientific discovery: you think you're looking for A, and you find B — and B is more interesting than A."

In unsupervised learning:

> "Good results appear by chance. The smart move is not to ignore these accidents, but to systematically identify, recycle, and amplify them. Chance contains necessity — if a certain kind of result keeps reappearing, it must reflect some deeper regularity."

RGSUD's philosophy: don't wait for perfect supervision. Find exploitable signals within the existing training dynamics.

6. Takeaways

If you're working on unsupervised/self-supervised learning, ask yourself:

1. "Are there 'lucky accidents' during training that can be exploited?" 2. "Can I design a reward function to identify high-quality outputs?" 3. "Can self-reinforcement replace external supervision?" 4. "Do the training dynamics themselves contain usable supervisory signals?"

RGSUD reminds us: unsupervised does not mean "no signal." The signal is in the training process — we just need clever methods to discover and use it.

In the wilderness of unsupervised learning, reward-guided self-reinforcement is a lamp — not sunlight from outside, but a torch you light yourself.