Automatic Enhancement of Low Light Level Image on the Basis of a Gaussian Mixture Model

Liju Yin, Tingdong Kou, Xuan Wang, Guofeng Zou, Jinfeng Pan, Zhongshang ZhuAsian 2019 · DOI

Abstract

Low-light imaging is a critical medium for transmitting information in environments with minimal illumination. However, the quality of these images is often degraded by external factors, resulting in blurry and low-contrast outputs. This paper proposes an automatic enhancement method based on a Gaussian Mixture Model (GMM). The approach models the image histogram as a GMM, solved via an accelerated Expectation-Maximization (EM) algorithm. By partitioning the histogram at the intersections of Gaussian clusters and establishing a mapping relationship for the output image, the algorithm identifies the optimal number of clusters and accelerates convergence. Objective evaluations using Laplace operator values, grayscale average gradients, and contrast metrics demonstrate that this method effectively improves image contrast while preserving fine details.

Introduction

Capturing clear images in low-light environments is essential for applications ranging from night-vision surveillance to medical diagnostics. In these settings, external noise often results in low-contrast, blurry outputs where targets are difficult to recognize.

Traditional solutions, such as global Histogram Equalization (HE), often fail because they over-stretch the grayscale range, leading to lost signals and amplified background noise. More advanced methods like Contrast Limited Adaptive Histogram Equalization (CLAHE) offer improvements but can still generate significant noise at grayscale boundaries or produce images that appear unnaturally bright or dark. This research introduces a GMM-based approach designed to enhance contrast and detail automatically while avoiding these "undesirable phenomena."

Methodology

The core of the proposed approach is treating the image's histogram not as a single block of data, but as a collection of overlapping distributions representing different intensity zones.

Gaussian Mixture Modeling

The image histogram is modeled as a sum of $N$ Gaussian clusters (typically between 2 and 6). Each cluster represents a specific range of grayscale intensities with its own mean and variance. To find the optimal parameters for these clusters quickly, the authors utilize an accelerated convergence EM algorithm based on robustness. This allows the system to identify the best number of clusters and converge on a solution in approximately 100 iterations.

Gaussian Mixture estimated by EM algorithm showing the partitioning of the image histogram into distinct clusters.

Histogram Partitioning and Mapping

Once the clusters are defined, the algorithm identifies the mathematical intersections between these Gaussian curves. These intersection points are used to divide the total dynamic range into sub-intervals. Finally, the gray values in each sub-interval are mapped to a new output range. This process effectively "stretches" the parts of the image that are hard for the human eye to distinguish, boosting local contrast while preserving the fine textures of the original scene.

Experimental Setup and Results

The researchers validated the algorithm using a custom-built micro-light imaging optical platform. The setup included a Multi-Pixel Photon Counter (MPPC) detector and a tungsten light source, specifically focusing on "avalanche excess noise," which is the primary interference in high-sensitivity semiconductor photodetectors.

Visual comparison of low-light image enhancement results across different methodologies.

The GMM-based method was compared against HE and CLAHE using both subjective visual analysis and objective metrics. The key findings include:

  • Contrast: The proposed method increased image contrast by nearly 28% over existing algorithms.
  • Edge Clarity: The Laplace operator value, which measures contour distinctness, improved by nearly 26%.
  • Texture Detail: The grayscale average gradient increased by 2.3%, indicating clearer texture information.

Qualitatively, the GMM method avoided the "washed out" look of HE and the boundary noise common in CLAHE. By maintaining the human visual characteristics of the scene, the output is significantly more suitable for target recognition tasks.

Conclusion

This research demonstrates that modeling image histograms with a Gaussian Mixture Model provides a robust framework for low-light enhancement. By partitioning the dynamic range based on cluster intersections, the algorithm achieves superior contrast and edge clarity without the artifacts associated with traditional histogram manipulation. Future work may involve adapting this approach for color (RGB) images and optimizing the EM algorithm for real-time processing in live night-vision applications.

Summary updated

Abstract

As the first medium to transmit information under a low light level environment. The low light level image is needed in hot-light image imaging technology. The quality of the image will be reduced given the influence of external factors. For example, a sampled image may become blurry. This paper proposes a method for automatic enhancement of low light level image on the basis of a Gaussian mixture model. First, the histogram of the image is modeled with a Gaussian mixture model that is solved by the expectation maximization algorithm of accelerated convergence. The histogram is then partitioned according to the intersection of each cluster. Finally, the mapping relationship of the cluster to which the output image belongs is ascertained and the final enhancement image is obtained. This algorithm can be used to identify the optimal number of clusters and accelerate the convergence speed of the algorithm. Objective evaluation of the Laplace operator value, as well as the grayscale average gradient and contrast (Tab. 1), indicates that the algorithm effectively improves image contrast while maintaining the details of the image.
BibTeX
@inproceedings{yin2019automatic,
  author = {Liju Yin and Tingdong Kou and Xuan Wang and Guofeng Zou and Jinfeng Pan and Zhongshang Zhu},
  title = {{Automatic Enhancement of Low Light Level Image on the Basis of a Gaussian Mixture Model}},
  booktitle = {Asian},
  year = {2019},
  doi = {10.24203/ajeel.v7i3.5818},
}
Figure