DarkIR: Robust Low-Light Image Restoration

Daniel Feijoo,Juan C. Benito,Alvaro Garcia,Marcos V. Conde

2024-12-18

Abstract:Photography during night or in dark conditions typically suffers from noise, low light and blurring issues due to the dim environment and the common use of long exposure. Although Deblurring and Low-light Image Enhancement (LLIE) are related under these conditions, most approaches in image restoration solve these tasks separately. In this paper, we present an efficient and robust neural network for multi-task low-light image restoration. Instead of following the current tendency of Transformer-based models, we propose new attention mechanisms to enhance the receptive field of efficient CNNs. Our method reduces the computational costs in terms of parameters and MAC operations compared to previous methods. Our model, DarkIR, achieves new state-of-the-art results on the popular LOLBlur, LOLv2 and Real-LOLBlur datasets, being able to generalize on real-world night and dark images. Code and models at <a class="link-external link-https" href="https://github.com/cidautai/DarkIR" rel="external noopener nofollow">this https URL</a>

Computer Vision and Pattern Recognition,Image and Video Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that images taken in low - light environments are usually affected by noise, blurring and insufficient illumination, especially at night or in dimly - lit situations. Specifically, the paper aims to solve the following problems:

1. **Joint processing of low - light image enhancement (LLIE) and deblurring**: Most existing methods handle low - light enhancement and deblurring tasks separately, which leads to high model complexity, large computational cost, and poor performance in practical applications. This paper proposes a multi - task neural network model that can handle these two problems simultaneously within the same framework.

2. **Image degradation problems in low - light conditions**: In low - light conditions, due to the limitations of camera sensors (such as motion blur and increased noise caused by long - exposure), the image quality drops significantly. This paper effectively solves these problems by introducing new attention mechanisms and frequency - domain processing methods.

3. **Improving model efficiency and generalization ability**: Although some existing deep - learning models have good performance, they have a large number of parameters and high computational complexity, and are difficult to be deployed on resource - limited devices. The DarkIR model proposed in this paper not only achieves the latest SOTA results on multiple datasets, but also has a significant reduction in the number of parameters and computational complexity, thereby improving the practicality and generalization ability of the model.

### Specific problem description

According to the mathematical formula in the paper, the formation process of a low - light image can be expressed as:
\[ y=\gamma(x\otimes k)+n \]
where:
- \( y \) is the observed low - light image;
- \( x \) is the original clear image;
- \( k \) is the lens point - spread - function (PSF) blur kernel;
- \( n \) is the sensor noise;
- \( \gamma \) is a function for controlling the dynamic range and pixel saturation;
- \( \otimes \) represents the convolution operation.

### Main contributions of the paper

1. **Designed a lightweight neural network**, which combines frequency - domain attention and large - receptive - field spatial attention, and can efficiently handle low - light enhancement and deblurring tasks.
2. **Achieved the latest SOTA results on datasets such as LOLBlur and Real - LOLBlur**, and has a lower computational cost.
3. **Proposed a new benchmark for multi - task night - scene / low - light image enhancement**, which can better handle various degradation problems of low - light images in practical application scenarios.

In summary, through innovative network architecture and efficient algorithm design, this paper successfully solves the joint processing problem of low - light image enhancement and deblurring, and improves the efficiency and generalization ability of the model.