Prompting Rain Off: Evolving Compact Dual Prompts for Continual De-Raining

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Single-image rain removal has advanced, but struggles with changing data distributions due to catastrophic forgetting.
• Continual learning is crucial for real-world applications where data evolves over time.

Purpose of the Study:

• To propose an efficient, rehearsal-free continual learning framework for low-level vision tasks, specifically single-image deraining.
• To address catastrophic forgetting in dynamic data scenarios.

Main Methods:

• Evolving Compact Dual Prompt Learning (EcoDPL) framework utilizing image and feature level prompts.
• P-Fuser module for adaptive prompt fusion with attention maps.
• Grad-Tuner for knowledge compression via dictionary learning.

Main Results:

• EcoDPL demonstrates superior performance compared to existing continual learning techniques on diverse rain removal datasets.
• The framework shows robustness and versatility, performing well on both changing and static data distributions.
• EcoDPL achieves efficient knowledge retention without increasing training time or parameters.

Conclusions:

• EcoDPL offers an effective solution for continual learning in deraining, mitigating catastrophic forgetting.
• The proposed prompt-based approach enhances flexibility and knowledge compression for learning new tasks.
• The method proves robust and versatile for various data conditions in low-level vision.