Authors: Wenglei Wu, Hansen Feng, Yalin Shi
Editor: Wenglei Wu
Created on Nov 24, 2024; last modified on Nov 24, 2024.
Note: this post is a summarized version of the corresponding research paper, including only the selected content.
We spend half of our daily time in darkness and human eyes can still recognize items in a low-light environment. However, it is not always the case for cameras capturing dark scenes, especially for state-of-the-art object detection algorithms, which may suffer from degraded performance.
In order to solve this problem, many low-light enhancement methods have been proposed in recent years. However, these methods focus on making images visually appealing. For more advanced tasks, such as object detection, they may not be as powerful as expected.
To address this issue, we proposed an end-to-end deep learning model that integrates low-light enhancement methods well into object detection algorithms.
| Methods | References |
|---|---|
| Histogram Equalization | / |
| Adaptive Histogram Equalization | 24 |
| Contrast Limited Adaptive Histogram Equalization | 23, 27 |
| Methods | References |
|---|---|
| Single-Scale Retinex | 10 |
| Multi-Scale Retinex | 9 |
| Multi-Scale Retinex with Color Restoration | 19 |
| Methods | References |
|---|---|
| Deep Auto-encoder Approach | 11 |
| MSR-net | 13 |
| Deep Single Image Contrast Enhancement | 7 |
| Retinex-Net | 2 |
| Data-driven Approach | 1 |
| EEMEFN | 21 |
| ZERO-DCE | 4 |
| Enlighten-GAN | 8 |
| SCI | 20 |
Two-stage detectors are a popular class of object detection models that typically consist of two main components: region proposal generation and object classification. These detectors follow a two-step approach, where the first stage generates a set of candidate object proposals, and the second stage performs classification and refinement on these proposals.