Rethinking the Importance of High-Frequency Components in Transformers for Image Restoration.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Transformer models excel at image restoration due to long-range dependency modeling.
• However, Transformers struggle with local information, limiting their ability to remove complex degradations.
• Existing methods often dilute crucial local patterns through self-attention mechanisms.

Purpose of the Study:

• To propose a novel Transformer-based approach, High-frequency Injected Transformer (HIT), for superior image restoration.
• To address the limitation of Transformers in capturing local image details.
• To enhance feature representation by incorporating high-frequency information and multi-scale aggregation.

Main Methods:

• Introduced a window-wise injection module (WIM) to integrate high-frequency details into feature maps.
• Developed a bidirectional interaction module (BIM) for multi-scale feature aggregation via mutual reinforcement.
• Incorporated a spatial enhancement unit (SEU) to preserve spatial relationships lost during channel-wise computations.

Main Results:

• HIT demonstrated superior performance across six diverse image restoration tasks: real noise, rain streak, blur, flare, underwater, and low-light conditions.
• The proposed method achieved favorable results compared to state-of-the-art techniques.
• HIT operates with linear computational complexity, offering an efficient solution.

Conclusions:

• HIT effectively overcomes the limitations of standard Transformers in image restoration by incorporating local details.
• The novel modules (WIM, BIM, SEU) contribute to improved feature representation and restoration quality.
• HIT presents a promising and efficient approach for various real-world image restoration challenges.