DARK: Denoising, Amplification, Restoration Kit
Zhuoheng Li, Yuheng Pan, Houcheng Yu, Zhiheng Zhang
TL;DR
DARK addresses the challenge of improving low-light images with a lightweight, real-time framework that combines a Retinex-based illumination estimator with context-aware enhancement. It introduces novel blocks (SRCB, SCB) and a selective kernel fusion mechanism (SKFF) to efficiently fuse multi-scale features, drawing inspiration from MIRNet-v2 and Retinexformer while maintaining a small parameter count of $187{,}123$. On the LoL dataset, it achieves a PSNR of $21.16$ and an SSIM of $0.767$ with about 80 minutes of training on a consumer GPU, demonstrating strong performance with low computational cost. The work highlights practical impact by enabling daily-use low-light enhancement on standard hardware, with plans to further improve denoising and generalization across additional datasets.
Abstract
This paper introduces a novel lightweight computational framework for enhancing images under low-light conditions, utilizing advanced machine learning and convolutional neural networks (CNNs). Traditional enhancement techniques often fail to adequately address issues like noise, color distortion, and detail loss in challenging lighting environments. Our approach leverages insights from the Retinex theory and recent advances in image restoration networks to develop a streamlined model that efficiently processes illumination components and integrates context-sensitive enhancements through optimized convolutional blocks. This results in significantly improved image clarity and color fidelity, while avoiding over-enhancement and unnatural color shifts. Crucially, our model is designed to be lightweight, ensuring low computational demand and suitability for real-time applications on standard consumer hardware. Performance evaluations confirm that our model not only surpasses existing methods in enhancing low-light images but also maintains a minimal computational footprint.