Learning Dual Transformers for All-in-One Image Restoration From a Frequency Perspective

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image restoration aims to recover degraded images using computational methods.
• Existing models often struggle with multiple degradation types, requiring specialized architectures.
• A unified approach for handling diverse image degradations is highly desirable.

Purpose of the Study:

• To develop a single, adaptable model for comprehensive image restoration tasks.
• To effectively extract and utilize degradation representations for guiding the restoration process.
• To improve the performance and versatility of image restoration models.

Main Methods:

• Proposed a dual-transformer approach: a frequency-aware degradation estimation transformer (Dformer) and a degradation-adaptive restoration transformer (Rformer).
• Dformer analyzes degradations across different frequency components to learn robust priors.
• Rformer utilizes a degradation-adaptive self-attention (DA-SA) module guided by learned representations.

Main Results:

• Outperformed existing methods on five key restoration tasks: denoising, deraining, dehazing, deblurring, and low-light enhancement.
• Demonstrated effectiveness in handling real-world, spatially variant, and unseen degradation levels.
• The frequency-aware approach enables better adaptation to specific degradation characteristics.

Conclusions:

• The proposed dual-transformer model offers a unified and effective solution for all-in-one image restoration.
• Frequency analysis is crucial for understanding and adapting to various image degradations.
• This method advances the state-of-the-art in handling diverse and complex image restoration challenges.