Robust contrast enhancement method using a retinex model with adaptive brightness for detection applications.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Low light image enhancement is crucial for machine cognition but challenging due to artifacts and data dependency in existing methods.
• Realistic frameworks must preserve brightness and contrast robustly, unlike direct methods that amplify artifacts or network-based methods requiring extensive paired datasets.
• Degraded visual conditions like extreme darkness, low light, and back-light pose significant hurdles for current image enhancement techniques.

Purpose of the Study:

• To present a new framework for deep image enhancement in degraded visual conditions, addressing limitations of existing methods.
• To develop a method that preserves brightness, color, and contrast without relying on paired training data.
• To improve the applicability of image enhancement for machine cognition tasks in real-world scenarios.

Main Methods:

• Utilizes retinex-based image decomposition to separate reflection and illumination components for independent enhancement.
• Employs a comprehensive weighting strategy to constrain decomposition and enhance contrast by disrupting high-frequency irregularities.
• Guides the illumination component with a high-frequency component for structure and texture preservation, introducing the deep into darkness network (D2D-Net).

Main Results:

• The proposed framework successfully enhances images in degraded visual conditions, preserving visual details with balanced brightness and contrast.
• D2D-Net maintains visual smoothness without compromising image quality, outperforming existing approaches.
• Experimental results demonstrate the method's superiority, particularly for object detection in extremely dark scenarios.

Conclusions:

• The proposed retinex-based decomposition and D2D-Net offer a robust solution for low light image enhancement, independent of training data type.
• The method effectively balances visual smoothness and image quality, making it suitable for advanced machine cognition applications.
• This framework significantly advances the viability of interactive visual applications in challenging, low-light environments.