Lightweight Low-Light Image Enhancement via Distribution-Normalizing Preprocessing and Depthwise U-Net
Summary: arXiv:2604.11071v2 Announce Type: replace-cross
Abstract: We present a lightweight two-stage framework for low-light image enhancement (LLIE) that achieves competitive perceptual quality with significantly fewer parameters than existing methods. Our approach combines frozen algorithm-based preprocessing with a compact U-Net built entirely from depthwise-separable convolutions. The preprocessing normalizes the input distribution by providing complementary brightness-corrected views, enabling the trainable network to focus on residual color correction. Our method achieved 4th place in the CVPR 2026 NTIRE Efficient Low-Light Image Enhancement Challenge. We further provide extended benchmarks and ablations to demonstrate the general effectiveness of our methods.
Introduction
Low-light image enhancement is a critical area in computer vision, aimed at improving the quality of images captured in insufficient lighting conditions. Traditional methods often struggle with heavy computational requirements and complex architectures, making them less suitable for real-time applications. In our latest research, we introduce a novel approach that addresses these challenges through a streamlined two-stage framework.
Methodology
Our proposed framework consists of two main components:
- Distribution-Normalizing Preprocessing: This initial stage employs a frozen algorithm to preprocess the input images. The preprocessing step is crucial as it normalizes the distribution of pixel intensity, offering brightness-corrected views of the images. This normalization allows the subsequent network to concentrate on the finer aspects of color correction rather than compensating for extreme lighting conditions.
- Depthwise U-Net Architecture: The second stage utilizes a compact U-Net architecture, constructed entirely from depthwise-separable convolutions. This design choice not only reduces the number of parameters significantly but also maintains competitive performance. The U-Net architecture is well-suited for image enhancement tasks due to its ability to capture contextual information efficiently.
Results
Our lightweight framework has demonstrated impressive results in various benchmarks. We participated in the CVPR 2026 NTIRE Efficient Low-Light Image Enhancement Challenge, where our method secured a commendable 4th place among numerous competing algorithms. The results underscore the effectiveness of our approach in terms of both efficiency and quality.
Extended Benchmarks and Ablations
To further validate our method, we conducted extensive ablation studies and comparisons with existing state-of-the-art techniques. These evaluations showcase the general effectiveness of our lightweight model across diverse datasets, highlighting its ability to deliver high-quality enhancements with minimal computational overhead.
Conclusion
In conclusion, our two-stage framework for low-light image enhancement sets a new standard for efficiency and quality. By leveraging distribution-normalizing preprocessing and a depthwise U-Net architecture, we provide a solution that is not only lightweight but also competitive against more complex models. This advancement opens new avenues for real-time applications in various fields, including photography, surveillance, and autonomous driving, where low-light conditions frequently pose challenges.
We look forward to future research endeavors aimed at further improving low-light image enhancement techniques and exploring their applications in real-world scenarios.