Deep Joint Unrolling for Deblurring and Low-Light Image Enhancement (JUDE).pdf

Tu Vo,Chan Y. Park

2024-12-10

Abstract:Low-light and blurring issues are prevalent when capturing photos at night, often due to the use of long exposure to address dim environments. Addressing these joint problems can be challenging and error-prone if an end-to-end model is trained without incorporating an appropriate physical model. In this paper, we introduce JUDE, a Deep Joint Unrolling for Deblurring and Low-Light Image Enhancement, inspired by the image physical model. Based on Retinex theory and the blurring model, the low-light blurry input is iteratively deblurred and decomposed, producing sharp low-light reflectance and illuminance through an unrolling mechanism. Additionally, we incorporate various modules to estimate the initial blur kernel, enhance brightness, and eliminate noise in the final image. Comprehensive experiments on LOL-Blur and Real-LOL-Blur demonstrate that our method outperforms existing techniques both quantitatively and qualitatively.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

This paper attempts to solve the problems of image blurring and insufficient light caused by long - exposure when taking photos in low - light conditions. Specifically, the paper proposes a deep joint unrolling network named JUDE (DeepJointUnrolling for Deblurring and Low - Light Image Enhancement), aiming to deal with low - light enhancement and image deblurring simultaneously.

### Problem Background
1. **Low - light and Blurring Problems**: Photos taken at night or in low - light environments usually require long - exposure to fully illuminate the scene. However, long - exposure is prone to cause motion blurring due to camera movement or dynamic elements in the frame. This makes night - scene photography often face the problems of insufficient light and blurring.
2. **Limitations of Existing Methods**:
- Early methods regarded low - light enhancement and image deblurring as two independent problems and processed them separately. This approach lacks a collaborative solution and cannot effectively improve the brightness and clarity of the image simultaneously.
- Most of the existing deblurring methods are trained based on daytime images and have poor performance in night - scene.
- Although end - to - end models can handle these two problems simultaneously, they are regarded as black - box models, which are difficult to analyze their behaviors, and rely on a large amount of data, limiting their generalization ability.

### Proposal of JUDE
To solve the above problems, the paper introduces JUDE, which is based on the Retinex theory and blurring model. Through an iterative unrolling mechanism, it gradually removes blurring and decomposes low - light images, and finally generates clear low - light reflectance and luminance maps. In addition, JUDE also combines multiple modules to estimate the initial blurring kernel, enhance luminance and eliminate noise.

### Main Contributions
1. **Innovative Deep Joint Unrolling Network**: JUDE contains four modules: blurring kernel estimation, optimization, luminance enhancement and reflectance denoising. These modules work together to solve the low - light and blurring problems.
2. **Deep Networking of the Optimization Module**: The optimization process is transformed into a deep network, taking advantage of the powerful modeling ability of the learning method to adaptively combine data - driven priors.
3. **Experimental Verification**: Experiments on the LOL - Blur and Real - LOL - Blur datasets show that JUDE outperforms existing methods in both quantitative and qualitative indicators, demonstrating its superior performance on synthetic and real - world datasets.

### Summary
By combining the advantages of physical models and deep learning, JUDE provides an effective solution that can handle low - light enhancement and image deblurring simultaneously, and has strong generalization ability and interpretability.