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

Forum topic · 小凯 · 2026-05-04

Summary

RGSUD (Reward-Guided Self-Reinforcement Unsupervised Deraining) is an unsupervised image deraining framework that addresses the lack of paired rainy/clean training data. Its key insight is that high-quality derained outputs occasionally emerge during training even without ground-truth supervision. The method operates in two stages: (1) Reward Recycling, where a reward function identifies high-quality outputs based on statistics of clean images (e.g., gradient distributions), semantic consistency, and naturalness; and (2) Self-Reinforcement, where these selected outputs serve as pseudo-targets for further training, creating a positive feedback loop of improving results. Unlike conventional unsupervised approaches that rely solely on static data-distribution assumptions—assumptions broken by rain's variable size, density, orientation, and coupling with background textures—RGSUD dynamically exploits training dynamics itself as a supervision signal, akin to intrinsic rewards in reinforcement learning. The paper is "Unpaired Image Deraining Using Reward-Guided Self-Reinforcement Strategy" by Yinghao Chen, Yeying Jin, Xiang Chen, Yanyan Wei, Ziyang Yan, and Yaowen Fu (arXiv: 2605.00719, 2026-04-30). The approach offers transferable lessons for unsupervised and self-supervised learning: identify emergent high-quality outputs, design reward functions around intrinsic image properties, and replace external supervision with self-reinforcement.

RGSUD: Unpaired Image Deraining with a Reward-Guided Self-Reinforcement Strategy

> 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

The Problem: Why Deraining Is Hard Without Pairs

Traditional deraining systems rely on paired "rainy → clean" images with supervised learning. In practice, however:

Real paired rainy/clean images are extremely difficult to collect
Synthetic rain often looks unrealistic
Real-world rain varies enormously in appearance

Unsupervised deraining avoids the need for paired data, but without strong constraints during training, models struggle to converge. Rain itself complicates matters: droplets vary in size, streaks differ in density and direction, and rain mixes with background textures—especially in dynamic scenes.

Without a "correct answer" to guide learning, unsupervised networks don't know what a good deraining result looks like, easily fall into local optima, and tend to produce blurry or over-smoothed outputs.

Key Insight

Even without paired supervision, high-quality derained results occasionally emerge during training. The question is how to recognize and exploit these "accidental wins."

The RGSUD Approach

RGSUD (Reward-Guided Self-Reinforcement Unsupervised Deraining) treats these occasional high-quality outputs as rewards and recycles them to guide further optimization, in two stages:

Stage 1 — Reward Recycling A reward function identifies good deraining results based on:

Statistical properties of rain-free images (e.g., gradient distributions)
Semantic consistency (main objects unchanged after deraining)
Naturalness (the output looks like a real rain-free image)

Stage 2 — Self-Reinforcement The identified high-quality outputs become pseudo-targets. The network learns to produce more results of this kind, forming a positive feedback loop: good results → more good results.

It's like a student solving problems without an answer key: when a solution turns out unusually elegant, they record it, analyze why it works, and deliberately repeat that quality.

Why Self-Rewarding Works

Conventional unsupervised methods assume the data distribution alone contains enough information. Rain's complexity breaks this assumption—rain distributions are coupled with the background, and purely statistical constraints fall short.

RGSUD's innovations:

Exploits training dynamics: rather than statically fitting the data distribution, it dynamically harvests emergent high-quality outputs during training
Self-discovery: the network discovers for itself what "good" means
No external supervision: rewards are computed from properties of the images themselves

This resembles intrinsic rewards in reinforcement learning—generated by the system itself rather than imposed externally.

Takeaways for Practitioners

If you work on unsupervised or self-supervised tasks, ask:

1. Are there "accidental wins" during training that can be exploited? 2. Can you design a reward function to recognize high-quality outputs? 3. Can self-reinforcement replace external supervision? 4. Does the training dynamics itself contain usable supervisory signals?

RGSUD is a reminder that unsupervised does not mean "no signal"—the signal lives inside the training process; we just need clever methods to find and use it. Reward-guided self-reinforcement is like a self-lit torch in the wilderness of unsupervised learning: not sunlight from outside, but a fire you light yourself.

*Source: arXiv:2605.00719, via zhichai.net*