CDIR: LoRA-Inspired Attention for Efficient Composite Degradation Image Restoration.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Specialized image restoration methods address single degradations.
• Real-world images often exhibit composite degradations.
• Existing unified models face high computational complexity.

Purpose of the Study:

• To propose an efficient attention module for composite degradation image restoration.
• To develop a unified restoration approach with reduced complexity.
• To enhance the adaptability of image restoration models.

Main Methods:

• Proposed an efficient attention module inspired by Low-Rank Adaptation (LoRA).
• Employed a dual-branch architecture with full and reduced resolution processing.
• Integrated dynamic operations guided by iteratively updated local and contextual priors.
• Introduced a multi-scale feed-forward network for efficiency.

Main Results:

• Achieved state-of-the-art performance on composite degradation benchmarks.
• Demonstrated significantly reduced computational complexity and fast inference speed.
• Showcased strong adaptability to various restoration tasks (dehazing, desnowing, deraining).

Conclusions:

• The proposed CDIR network offers an efficient and versatile solution for composite image restoration.
• CDIR is applicable to diverse domains including UHD, remote sensing, and medical imaging.
• The model's design balances performance with computational efficiency.