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' by Yinghao Chen et al. Unsupervised image deraining avoids the need for hard-to-obtain paired rainy/clean images, but lacks the strong constraints needed for stable convergence, often producing blurry or over-smoothed outputs. RGSUD exploits a key observation: even without supervision, high-quality deraining results occasionally emerge during training. It works in two stages: (1) Reward Recycling, where a reward function—based on clean-image statistics such as gradient distributions, semantic consistency, and naturalness—identifies these high-quality outputs; and (2) Self-Reinforcement, where the identified outputs serve as pseudo-targets, creating a positive feedback loop of good results guiding further optimization. The author frames this as an 'intrinsic reward' mechanism analogous to reinforcement learning, arguing that unsupervised learning still contains exploitable signals within its training dynamics. The post also discusses why self-rewards work, connects the idea to serendipitous scientific discovery, and offers practical questions for anyone working on unsupervised or self-supervised 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

The "Rain Gets Heavier" Problem for AI

Imagine an AI system tasked with removing rain streaks from photos. The traditional approach is supervised learning: collect paired rainy/clean images and train the model to map one to the other. But in reality, real paired data is nearly impossible to obtain, synthetic rain is unrealistic, and real-world rain varies endlessly.

Unsupervised deraining sounds ideal—no paired data required. The catch: without strong constraints during training, models struggle to converge.

Why Unsupervised Deraining Is Hard

Rain is complicated: droplets vary in size, streak density and direction differ, rain mixes with background textures, and dynamic scenes make things worse.

The unsupervised dilemma:

There is no "correct answer" to guide learning
The network doesn't know what a good deraining result looks like
It easily falls into local optima, producing blurry or over-smoothed outputs

Key insight: even without paired supervision, high-quality deraining results occasionally appear during training. The question is how to identify and exploit these "happy accidents."

RGSUD: Reward-Guided Self-Reinforcement

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

> Treat the high-quality results that偶然 emerge during training as "rewards," and recycle them to guide subsequent optimization.

Two-stage strategy:

Stage 1: Reward Recycling

During training, the network occasionally produces good deraining results
A reward function identifies these 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 serve as pseudo-targets
The network learns to produce more results of similar quality
A positive feedback loop: good results → more good results

It's like a student solving problems without an answer key: occasionally they discover an especially elegant solution, record why it works, and deliberately reproduce that "goodness."

Why "Self-Reward" Works

Traditional unsupervised methods assume the data distribution contains enough information for the network to learn from statistics alone. Rain's complexity breaks this assumption—rain distributions are entangled with the background, so purely statistical constraints fall short.

RGSUD's innovations:

Exploiting training dynamics: instead of statically learning the data distribution, it dynamically leverages emergent "happy accidents" during training
Self-discovery: the network discovers for itself what "good" looks like
No external supervision: the reward function relies only on properties of the images themselves

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

Lessons to Take Away

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

1. Are there "happy 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 lives inside the training process—we just need clever methods to find and use it.

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