CODEN: combined optimization-based decomposition and learning-based enhancement network for Retinex-based brightness...

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Low-light conditions significantly degrade image quality, impacting visibility and detail.
• Existing low-light enhancement methods often struggle with preserving details or achieving balanced brightness and contrast.

Purpose of the Study:

• To develop a novel, end-to-end trainable method for simultaneously enhancing brightness and contrast in low-light images.
• To improve upon state-of-the-art low-light image enhancement techniques.

Main Methods:

• A two-step approach combining Retinex-based decomposition and a deep illumination enhancement network (IEN).
• The first step decomposes images into illumination and reflectance components using model-based optimization and learning for edge-preserved smoothing and detail-preserved denoising.
• The second step utilizes an IEN with residual squeeze and excitation blocks (RSEBs) for illumination enhancement.

Main Results:

• The proposed method successfully enhances both brightness and contrast in low-light images.
• Experimental results demonstrate superior performance over existing state-of-the-art methods.
• Objective and subjective evaluations confirm the effectiveness of the approach.

Conclusions:

• The combined optimization-based decomposition and enhancement network offers a robust solution for low-light image enhancement.
• The end-to-end trainable nature and effective component separation contribute to improved image quality.
• This method provides a significant advancement in enhancing visual fidelity for underexposed imagery.