A Zero-Shot Low Light Image Enhancement Method Integrating Gating Mechanism.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Harsh ambient lighting causes image degradation like low brightness and noise.
• Deep learning methods exist but require paired training data, which is often unavailable.
• Existing methods struggle with the challenge of acquiring paired training data for low-light image enhancement.

Purpose of the Study:

• To propose a zero-reference image enhancement network for low-light conditions.
• To address the limitation of paired training data in existing deep learning approaches.
• To develop an unsupervised method for improving illumination in underexposed images.

Main Methods:

• Utilizes an improved Encoder-Decoder structure for feature extraction and parameter matrix generation.
• Constructs an enhancement curve using the parameter matrix for iterative image enhancement.
• Employs four non-reference loss functions for unsupervised training of the parameter estimation network.

Main Results:

• Achieved superior performance over existing methods on NIQE, PIQE, and BRISQUE non-reference evaluation indices.
• Ablation experiments confirmed the effectiveness of key components within the proposed network.
• Demonstrated practical feasibility and performance on both PC and mobile devices.

Conclusions:

• The proposed zero-reference network effectively enhances low-light images without paired data.
• The unsupervised approach with non-reference loss functions proves robust and efficient.
• The method is suitable for practical applications, offering improved image quality and performance across devices.