中文 |

Newsroom

Innovative Zero-reference Method Boosts Low-light Image Quality

Sep 03, 2024

Researchers from the Changchun Institute of Optics, Fine Mechanics, and Physics of the Chinese Academy of Sciences recently developed a novel low-light image enhancement (LLIE) method, the Zero-Reference Camera Response Network (ZRCRN), which addresses limitations of existing LLIE methods. The study was published in Sensors.

Low-light images are prevalent in many fields. However, insufficient light intensity often results in images with low brightness, poor contrast, and high noise levels, hindering further processing and analysis. Traditional methods for LLIE, including histogram equalization and Retinex theory, either have poor enhancement effects or require complex parameter tuning. Besides, deep learning-based methods often require large amounts of labeled data, limiting their adaptability.

To address these issues, researchers proposed the ZRCRN, a fast and efficient method that leverages a camera response model. The key innovation lies in the establishment of a double-layer parameter-generating network, which automatically extracts the exposure ratio (K) from the radiation map obtained by inverting the input image through a camera response function. This exposure ratio serves as the parameter for a brightness transformation function, transforming the low-light image into an enhanced version in a single step.

Furthermore, researchers designed two reference-free loss functions, i.e., a contrast-preserving brightness loss and an edge-preserving smoothness loss. The former ensures that the brightness distribution in the original image is retained while enhancing contrast, while the latter promotes noise reduction and detail enhancement. These loss functions, without requiring paired reference images, significantly improve the generalization ability of the model.

Extensive experiments were conducted on several standard LLIE datasets and the DARK FACE face detection dataset. The results demonstrated that the ZRCRN achieved superior performance both subjectively and objectively, with an enhancement speed more than twice that of similar methods.

The ZRCRN is an efficient method to LLIE, and achieves fast and accurate enhancement, paving the way for advanced machine vision applications in challenging lighting conditions. It holds the potential for various applications. By enabling real-time and high-quality enhancement of low-light images, the method can improve the performance of autonomous vehicles and remote sensing platforms. Its zero-reference nature makes it highly adaptable to different scenarios, reducing the dependence on large labeled datasets.

Contact

NIE Ting

Changchun lnstitute of Optics, Fine Mechanics and Physics

E-mail:

Fast, Zero-Reference Low-Light Image Enhancement with Camera Response Model

Related Articles
Contact Us
  • 86-10-68597521 (day)

    86-10-68597289 (night)

  • 86-10-68511095 (day)

    86-10-68512458 (night)

  • cas_en@cas.cn

  • 52 Sanlihe Rd., Xicheng District,

    Beijing, China (100864)

Copyright © 2002 - Chinese Academy of Sciences