Unsupervised Low-Light Image Enhancement Based on Generative Adversarial Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Low-light conditions degrade image quality, impacting visual perception and downstream tasks.
• Existing low-light image enhancement methods often struggle with detail recovery and artifact generation.

Purpose of the Study:

• To propose a novel generative adversarial network (GAN) for superior low-light image enhancement.
• To enhance the perceptual quality and recover fine details in images captured under challenging lighting.

Main Methods:

• A generator architecture incorporating residual modules, hybrid attention, and parallel dilated convolutions to capture multi-scale features.
• Utilized skip connections for effective fusion of shallow and deep features, preventing information loss.
• A discriminator designed for improved discrimination ability and an enhanced loss function with pixel loss for detailed information recovery.

Main Results:

• The proposed GAN method demonstrated superior performance in enhancing low-light images.
• Qualitative and quantitative evaluations showed significant improvements over seven other state-of-the-art methods.
• The method effectively recovered detailed information and improved overall image perceptual quality.

Conclusions:

• The novel GAN architecture with its specific modules and loss function offers a powerful solution for low-light image enhancement.
• The proposed approach effectively addresses limitations of existing methods, providing enhanced visual quality.
• This work contributes a robust method for improving images captured in low-light environments.