DarkIR: Robust Low-Light Image Restoration

TL;DR

DarkIR tackles the challenge of restoring night/dark images degraded by noise, blur, and limited illumination by proposing a joint low-light enhancement and deblurring framework. It introduces a lightweight CNN with a Fourier-domain encoder for illumination correction and a spatial, large-receptive-field decoder for deblurring, inspired by Metaformer blocks and augmented with Di-SpAM. The loss combines pixel-level, perceptual, edge, and architecture-guiding components to balance fidelity and visual quality, achieving state-of-the-art results on LOLBlur, Real-LOLBlur, and LOLv2 while reducing parameters and MACs compared to prior methods. The approach generalizes to real-world night scenes and real data, and the authors provide code and models to support reproducibility and integration into mobile and edge devices.

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 https://github.com/cidautai/DarkIR

Figures (14)

• Figure 1: Previous Low-light Image Enhancement (LLIE) and restoration methods are not robust to blur and illumination changes. Our multi-task model is able to restore real low-light images under varying illumination, noise and blur conditions. Zoom-in to see details.
• Figure 2: General diagram of DarkIR. The neural network follows an encoder-decoder architecture. We use different blocks for encoding and decoding that follow the Metaformer structure yu2022metaformer. The encoder focuses on the low-light illumination issues using Fourier information. Thus, the encoder produces a low-resolution reconstructed image $\hat{x}_{\downarrow8}$ with corrected illumination. The decoder focuses on upscaling and reducing the blur using the prior illumination-enhanced encoded features. To achieve this, the decoder uses large receptive field spatial attention. This design allows our lightweight model to have less parameters and FLOPs than previous methods.
• Figure 3: Qualitative comparisons on the LOLBlur dataset (synthetic samples from the testset).
• Figure 4: Additional visual comparisons on the LOLBlurlednet dataset with 2-step pipelines. DarkIR generates much sharper images with visually pleasing results. (Zoom in for best view).
• Figure 5: Qualitative comparison on real night scenes from the RealBlurLOL dataset.