A Two-Stage Network for Zero-Shot Low-Illumination Image Restoration.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-illumination images suffer from noise, artifacts, and darkening due to poor lighting and equipment limitations.
• These image quality issues negatively impact high-level image understanding tasks.
• Existing methods often struggle to balance noise reduction with detail preservation.

Purpose of the Study:

• To develop an effective method for restoring low-illumination images.
• To improve the visual quality and utility of images captured in poor lighting conditions.
• To enhance performance in subsequent image analysis tasks.

Main Methods:

• A two-stage network approach is proposed, comprising a Decom-Net and an Enhance-Net.
• The Decom-Net decomposes low-illumination images into reflectance, illumination, and feature maps.
• Noise is suppressed in reflectance and illumination maps, while feature maps retain image details. The Enhance-Net adjusts illumination, and the maps are fused for the final output.
• The network is optimized using a novel loss function in a zero-shot manner.

Main Results:

• The proposed network effectively restores low-illumination images.
• Experimental results show superior performance compared to existing methods.
• Both objective evaluation metrics and visual quality are significantly improved.

Conclusions:

• The proposed two-stage network offers a robust solution for low-illumination image enhancement.
• Decomposition into multiple maps allows for targeted noise reduction and detail preservation.
• The method demonstrates potential for various computer vision applications requiring high-quality image input.