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

WACV'25

¹KC Machine Learning Lab

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.

Results

Synthesize Data

Scene 60 from the LOLBlur Dataset

JUDE

Input

LEDNet

FELI

FFTFormer -> RetinexFormer

FFTFormer

FourLLE -> FFTFormer

LLFormer -> FFTFormer

MIMO -> RetinexFormer

RetinexFormer -> FFTFormer

RetinexFormer

Scene 118 from the LOLBlur Dataset

JUDE

Input

LEDNet

FELI

FFTFormer -> RetinexFormer

FFTFormer

FourLLE -> FFTFormer

LLFormer -> FFTFormer

MIMO -> RetinexFormer

RetinexFormer -> FFTFormer

RetinexFormer

Real Data

Scene 205 from the Real-Blur Dataset

Input

JUDE

LEDNet

FELI

FFTFormer -> RetinexFormer

FFTFormer

FourLLE -> FFTFormer

LLFormer -> FFTFormer

MIMO -> RetinexFormer

RetinexFormer -> FFTFormer

RetinexFormer

Scene C0326 from the Real-Blur Dataset

Input

JUDE

LEDNet

FELI

FFTFormer -> RetinexFormer

FFTFormer

FourLLE -> FFTFormer

LLFormer -> FFTFormer

MIMO -> RetinexFormer

RetinexFormer -> FFTFormer

RetinexFormer

Performance Metrics

Benchmarking the LOL-Blur Dataset.

Model Name	PSNR ↑	SSIM ↑	LPIPS ↓
FourLLIE → FFTFormer	18.433	0.705	0.305
LLFormer → FFTFormer	20.290	0.792	0.212
RetinexFormer → FFTFormer	16.452	0.702	0.324
MIMO → RetinexFormer	17.024	0.770	0.271
FFTFormer → RetinexFormer	16.712	0.728	0.325
FFTFormer	19.889	0.858	0.139
RetinexFormer	25.505	0.862	0.240
LEDNet	25.740	0.850	0.224
FELI	26.728	0.914	0.132
JUDE	26.884	0.932	0.127

Benchmarking the Real-Blur Dataset.

Model Name	ARNIQA ↑	CONTRIQUE ↑	LIQE ↑	MUSIQ ↑	CLIPIQA ↑	DBCNN ↑
FourLLIE → FFTFormer	0.307	46.823	1.113	30.840	0.217	0.261
LLFormer → FFTFormer	0.401	44.743	1.158	36.534	0.208	0.257
RetinexFormer → FFTFormer	0.364	41.495	1.075	34.793	0.227	0.279
MIMO → RetinexFormer	0.413	40.773	1.137	33.242	0.207	0.276
FFTFormer → RetinexFormer	0.405	48.814	1.195	35.511	0.221	0.303
FFTFormer	0.402	38.005	1.141	32.079	0.289	0.307
RetinexFormer	0.418	43.410	1.074	31.782	0.187	0.232
LEDNet	0.419	49.828	1.414	43.623	0.281	0.306
FELI	0.429	42.354	1.155	33.669	0.207	0.239
JUDE	0.437	50.207	1.454	44.732	0.299	0.313

Citation

@article{tvo_jude,
  author    = {Tu Vo and Chan Y. Park},
  title     = {Deep Joint Unrolling for Deblurring and Low-Light Image Enhancement (JUDE)},
  booktitle = {The IEEE/CVF Winter Conference on Applications of Computer Vision},
  year      = {2025}
}