<aside> 👏
Key words:
<aside> 🧐
Claim:
This project is a reproduction of this paper, available at the following link:
LIME: Low-Light Image Enhancement via Illumination Map Estimation
Detailed project report can be found:
GitHub project link: </aside>
<aside> 📌
Project overview:
</aside>
Low-light images (Figure 1) often suffer from poor visibility, and many details are hidden in the dark area, which makes enhancement essential to improve their quality.
Current methods, such as Gamma correction, Histogram equalization, have been employed to enhance the low-light image. However, there still exist some limitation of these methods, which have been shown in the comparative examples in next Section Results.
The method proposed in the paper is based on Retinex theory, whose dominant assumption is that the color image can be decomposed into two factors, say reflectance and illumination. Reflectance mainly contains the color information of the image, when illumination mainly contains the contour (or kind of image’s intensity). In this paper, the key idea is to best estimate a great illumination map that can be used to enhance the image.
Figure 1: Low-light images
Here are examples which contain the original images and images enhanced by Histogram equalization (HE), Gamma correction (GC), and the method proposed in reference (LIME).
Figure 2: comparative results
🎆LIME : The image enhanced by LIME not only preserves a wealth of details but also produces enhanced visuals that are more aligned with human aesthetic preferences.
🎆HE : HE tends to produce overexposed results easily.
🎆GC : GC results in an optimized output with darker tones.
The project has been summarized as a poster for class exhibition.
In Retinex model, the low-light image of size $m\times n$ is interpreted as
$$ \mathbf{L} = \mathbf{R}~ \circ ~\mathbf{T}, \\ \text{} \\ \mathbf{L},~\mathbf{R},~\mathbf{T}\in\mathbb{R}^{m\times n \times3}, $$
$$ \begin{align*} &\mathbf{L}: \text{low-light image, matrix,} \\ &\mathbf{R}: \text{desired recovery image matrix},\\ &\mathbf{T}: \text{illumination map},\\ &\circ : \text{element-wise multiplication}. \end{align*} $$
For RGB images, the image is stored in matrix of size $m\times n \times 3$. In this project, one illumination map $\mathbf{T}$ is applied to 3 channels equivalently. Therefore, the primary objective in the following sections is to accurately estimate $\mathbf{T}$.