Learning Prompt Adapters for Forgetting-Free Continual Image Super-Resolution.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Continual image super-resolution (CISR) aims to adapt models to new tasks without forgetting previous ones.
• Existing prompt-based methods struggle with pixel-level restoration and domain discrimination in super-resolution.
• Catastrophic forgetting and the need for high adaptability are key challenges in CISR.

Purpose of the Study:

• To propose Learning Prompt Adapters (LPA) for effective CISR.
• To enhance fine-grained detail restoration and model adaptability in super-resolution.
• To preserve knowledge from previously learned tasks during continual learning.

Main Methods:

• Dynamically generating pixel-wise prompts using multi-granularity prompt bases and identities.
• Integrating adaptive prompts into a Transformer architecture.
• Organizing low-rank prompt bases with specific identities to manage cross-task differences.

Main Results:

• LPA significantly improves performance on fine-grained details in super-resolution tasks.
• The method enhances model adaptability to new tasks and preserves knowledge from prior tasks.
• Experiments on diverse datasets (NYU, RealSR, DIV2K, REDS, MANGA109) show superior performance over existing methods.

Conclusions:

• Learning Prompt Adapters offer an effective solution for continual image super-resolution.
• LPA addresses challenges of catastrophic forgetting and adaptability in low-level vision tasks.
• The proposed method demonstrates strong performance across various datasets and degradation types.