LGIT: local-global interaction transformer for low-light image denoising.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Transformer-based methods excel in visual tasks by capturing global image dependencies.
• Existing Transformers struggle with large, low-light noisy images due to high computational complexity and suboptimal spatial correlation optimization.

Purpose of the Study:

• To propose a novel Transformer architecture, the Local-Global Interaction Transformer (LGIT), for efficient and effective image denoising.
• To address the limitations of existing methods in handling large, noisy, low-light images.

Main Methods:

• Developed LGIT with an adaptive strategy for selecting relevant patches for global interaction, reducing computational load.
• Introduced a Top-N Patch Cross-Attention (TPCA) model guided by superpixel segmentation for enhanced nonlocal self-similarity utilization.
• Incorporated a Mixed-Scale Dual-Gated Feedforward Network (MDGFF) for multiscale local correlation extraction.

Main Results:

• LGIT demonstrated significantly improved qualitative and quantitative results on real-world image denoising datasets.
• The proposed TPCA and MDGFF components effectively improved information aggregation and local feature extraction.
• LGIT outperformed both state-of-the-art Convolutional Neural Network (CNN) and Transformer-based denoising methods.

Conclusions:

• LGIT offers a computationally efficient and highly effective solution for denoising large, noisy images under low-light conditions.
• The novel architecture successfully integrates local and global information processing for superior image restoration.
• LGIT represents a significant advancement in Transformer-based image denoising techniques.