Abstract
Low-light image enhancement (LLIE) requires a delicate balance between global illumination adjustment and the preservation of local structural details. Traditional methods often struggle with either the computational complexity of global modeling or the loss of fine-grained textures. This paper introduces WalMaFa (Wavelet-based Mamba with Fourier Adjustment), a novel framework that leverages the strengths of dual frequency domains. By utilizing Wavelet transforms for global brightness management and Fourier transforms for local texture refinement—integrated within an efficient State Space Model (Mamba) architecture—WalMaFa achieves state-of-the-art performance. Experimental results demonstrate that WalMaFa outperforms existing Transformer-based models in both visual quality and computational efficiency, reducing parameter counts and FLOPs significantly.
1. Introduction
Low-light image enhancement is a fundamental challenge in computer vision, aimed at recovering visibility in scenes with poor illumination, high noise, and color distortion. Current deep learning solutions primarily utilize Convolutional Neural Networks (CNNs) or Transformers. While CNNs are adept at capturing local features, they often fail to model the long-range dependencies necessary for uniform global brightness adjustment. Conversely, Transformers excel at global context modeling but suffer from quadratic computational complexity, making them impractical for high-resolution image processing.
Frequency-domain methods have emerged as a potential solution, yet many existing approaches treat frequency components as generic alternatives to spatial pixels. WalMaFa addresses this by identifying the specific roles of different frequency domains: Wavelet low-frequency components are optimal for global luminance, while Fourier phase components are superior for capturing edge and texture details.
2. Methodology
WalMaFa adopts an Encoder-Latent-Decoder architecture designed to exploit the "Global-Local-Global" characteristics of image frequency. The framework introduces two core modules:
2.1 Wavelet-based Mamba Block (WMB)
Integrated into the encoder and decoder stages, the WMB utilizes the Discrete Wavelet Transform (DWT) to decompose the image into sub-bands. The model applies a Channel-wise Mamba module specifically to the low-frequency (LL) component. Mamba, a State Space Model (SSM), provides the long-range modeling capabilities of Transformers but with linear complexity. By processing the LL component in the channel dimension, the WMB effectively models global lighting variations across the entire image without the heavy memory overhead of self-attention.
2.2 Fast Fourier Adjustment Block (FFAB)
Located in the latent stage of the network, the FFAB employs the Fast Fourier Transform (FFT) to refine the image. Unlike the WMB, which focuses on brightness, the FFAB targets the phase component of the Fourier transform. The authors' empirical analysis suggests that the phase component is more sensitive to local textures and structural edges than high-frequency Wavelet bands. This block acts as a high-precision tuner, sharpening details after the primary illumination adjustment has been performed.
3. Experiments and Results
3.1 Experimental Setup
The model was evaluated on several benchmark datasets, including LOL-v1 and LOL-v2 (Real & Synthetic). For qualitative assessment in real-world scenarios, unreferenced datasets such as DICM, LIME, and NPE were utilized. Performance was measured using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).
3.2 Quantitative Performance
WalMaFa demonstrates superior performance across standard benchmarks. On the LOL-v2-synthetic dataset, the model achieved a PSNR of 25.56 dB, notably higher than LLFormer (24.42 dB) and RetinexFormer (25.10 dB). This indicates a higher fidelity in both color reconstruction and noise suppression.
3.3 Computational Efficiency
A key advantage of the Mamba-based architecture is its efficiency. Compared to the popular LLFormer, WalMaFa requires 13.43M fewer parameters and reduces computational costs by approximately 22.6G FLOPs. Furthermore, inference speed is improved, with WalMaFa processing images in 0.068s compared to 0.077s for LLFormer, making it more suitable for real-time applications.
4. Discussion and Key Insights
The success of WalMaFa stems from the "Aha" moment regarding frequency domain roles. The researchers conducted a "swap" experiment, discovering that exchanging Wavelet LL components between dark and bright images yielded a significantly higher SSIM improvement (0.08dB) than swapping Fourier amplitude components. This confirmed that Wavelets are the superior tool for global brightness.
Additionally, by shifting from spatial scanning to channel-wise scanning within the Mamba blocks, the model more effectively captures the variability of color and brightness across different feature maps, leading to more natural color reproduction as confirmed by a 15-participant user study.
5. Conclusion and Future Work
WalMaFa represents a significant step forward in efficient low-light enhancement by strategically combining Wavelet and Fourier transforms with State Space Models. While the model shows great promise, the authors note that performance on very small datasets (like LOL-v1) can be limited by overfitting to global patterns. Future research will explore multi-directional spatial scanning in Mamba and the application of these frequency-domain techniques in alternative color spaces, such as HSV, to further decouple brightness from color information.
Summary updated