Unveiling Advanced Frequency Disentanglement Paradigm for Low-Light Image Enhancement

TL;DR

The paper addresses LLIE by decoupling frequency components rather than pursuing increasingly complex single architectures. It introduces a Laplace-pyramid–based frequency disentanglement framework with two modules: ACCA for coarse low-frequency adjustment and LDRM for Laplace-based high-frequency restoration, supervised by a low-frequency consistency loss $L_i$ so the stages can be trained decoupledly. The approach is compatible with CNNs, Transformers, flow-based, and diffusion LLIE models and yields up to $7.68$ dB PSNR gains with only about $8.8\times 10^4$ extra parameters, across five benchmarks. Ablation studies validate the necessity of the decoupled design and the pyramid-based restoration, while experiments show broad, consistent improvements with minimal computational overhead.

Abstract

Previous low-light image enhancement (LLIE) approaches, while employing frequency decomposition techniques to address the intertwined challenges of low frequency (e.g., illumination recovery) and high frequency (e.g., noise reduction), primarily focused on the development of dedicated and complex networks to achieve improved performance. In contrast, we reveal that an advanced disentanglement paradigm is sufficient to consistently enhance state-of-the-art methods with minimal computational overhead. Leveraging the image Laplace decomposition scheme, we propose a novel low-frequency consistency method, facilitating improved frequency disentanglement optimization. Our method, seamlessly integrating with various models such as CNNs, Transformers, and flow-based and diffusion models, demonstrates remarkable adaptability. Noteworthy improvements are showcased across five popular benchmarks, with up to 7.68dB gains on PSNR achieved for six state-of-the-art models. Impressively, our approach maintains efficiency with only 88K extra parameters, setting a new standard in the challenging realm of low-light image enhancement.

Figures (4)

• Figure 1: (a) Illustration of predictions and corresponding frequency-disentangled components. 'SNR-De' denote the improved version of SNR xu2022snr by our method. Visualization of the decoupled components shows that our method accurately recovers intensity in the low-frequency domain and effectively denoises in the high-frequency domain (see supplementary materials for more visual results). (b) Comparison between SOTA models and their improved versions on four representative benchmarks. It can be seen that our method significantly improves six representative SOTA models. $\bullet$ and $\star$ denote these baseline (SOTA) models and their enhanced version by our method, respectively.
• Figure 2: Overview of our proposed frequency disentanglement learning framework. It consists of two phases: (1) coarse phase: ACCA conducts coarse adjustment to initially enhance the input image $I$ and produce the preliminary result $I_l$, and (2) coarse-to-fine phase: LDRM integrates Laplace representations ($M, M_l$ from the $I, I_l$) for subsequent fine-grained restoration. Additionally, a low-frequency consistent loss ${\color{red}L_i}$ (Eq. \ref{['eq:loss']}) between the two phases is introduced to achieve effective disentanglement optimization.
• Figure 3: Qualitative evaluation on several benchmarks. Our integration provides accurate outcomes for both high-frequency (clearer image detail restoration) and low-frequency (more accurate illumination recovery) areas.
• Figure 4: (a) Visual validation of our disentanglement learning. (b-c) Impact of adopting different settings in our method: (b.1-2) different pyramid levels ($K$) in Laplace representation; (c.1-2) different $\alpha$ values to conduct low-frequency consistent supervision.