Low-light enhancement method with dual branch feature fusion and learnable regularized attention

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light imaging conditions introduce significant challenges such as color aberration and noise.
• These image quality issues hinder the performance of subsequent vision-based applications.
• Developing effective low-light image enhancement techniques is crucial for practical computer vision tasks.

Purpose of the Study:

• To propose a novel two-stage, size-controllable low-light image enhancement method.
• To address limitations in image quality caused by poor lighting conditions.
• To improve the utility of images captured in low-light environments for computer vision.

Main Methods:

• A Dual Fusion Enhancement Net (DFEN) employing a double U-Net architecture.
• Implementation of a dual branch feature fusion module for enhanced feature extraction and aggregation.
• Integration of a learnable regularized attention module and a cosine training strategy for balanced enhancement and smooth training transitions.

Main Results:

• DFEN demonstrates superior low-light image enhancement compared to existing methods.
• The algorithm achieves high performance with comparable model parameters.
• The lightest DFEN model achieves a processing speed of 11 FPS for 1224x1024 images on an RTX 3090 GPU.

Conclusions:

• The proposed DFEN effectively enhances low-light images by addressing brightness and detail deficiencies.
• DFEN offers a robust and efficient solution for improving image quality in challenging lighting conditions.
• The method shows significant potential for real-world applications requiring high-quality visual data from low-light environments.