SwinLightGAN a study of low-light image enhancement algorithms using depth residuals and transformer techniques

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing low-light image enhancement algorithms often sacrifice image details for improved brightness and contrast.
• There is a need for advanced methods that can effectively restore intricate details in images captured under suboptimal lighting.

Purpose of the Study:

• To introduce SwinLightGAN, a novel generative adversarial network (GAN) designed for effective low-light image enhancement with a focus on detail preservation.
• To leverage Swin Transformer technology for capturing multi-scale spatial features and global contexts in image enhancement.

Main Methods:

• The proposed SwinLightGAN integrates a Residual Jumping U-Net generator for local detail extraction and an illumination network enhanced by Swin Transformer.
• Adversarial training with multi-scale discriminators and a combination of loss functions were employed to ensure high-quality image generation.
• The network was trained and evaluated on multiple unpaired datasets to demonstrate its robustness and generalizability.

Main Results:

• SwinLightGAN successfully enhances images, producing results comparable to those taken in normal lighting conditions while preserving intricate details.
• Quantitative evaluation showed excellent performance with Naturalness Image Quality Evaluator (NIQE) scores between 5.193-5.397, Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) scores from 28.879-32.040, and Patch-based Image Quality Evaluator (PIQE) scores from 38.280-44.479.
• The experimental analysis confirmed the system's efficacy in delivering high-quality, detailed enhancements across diverse metrics.

Conclusions:

• SwinLightGAN represents a significant advancement in low-light image enhancement, particularly in its ability to restore and preserve fine image details.
• The integration of Swin Transformer technology within a GAN framework offers a powerful approach for capturing complex spatial and global image features.
• The study validates SwinLightGAN's superior performance and potential for real-world applications requiring high-fidelity image restoration in challenging lighting conditions.