From Darkness to Detail: Frequency-Aware SSMs for Low-Light Vision

TL;DR

This work addresses the need for real-time, high-quality low-light image enhancement on edge devices. It introduces ExpoMamba, a frequency-aware state-space model integrated into a U-Net, which decouples amplitude and phase processing via a Frequency State Space Block (FSSB) to improve brightness and structural fidelity. Key contributions include dual-branch amplitude/phase modeling, a latent-space color correction matrix, HDR feature-recovery gating, dynamic patch training, and a thorough evaluation showing strong PSNR/SSIM gains and superior efficiency versus transformer-based LLIE methods. The approach demonstrates clear practical impact for downstream vision tasks in low-light conditions, such as object detection and segmentation, while maintaining real-time performance on resource-constrained hardware.

Abstract

Low-light image enhancement remains a persistent challenge in computer vision, where state-of-the-art models are often hampered by hardware constraints and computational inefficiency, particularly at high resolutions. While foundational architectures like transformers and diffusion models have advanced the field, their computational complexity limits their deployment on edge devices. We introduce ExpoMamba, a novel architecture that integrates a frequency-aware state-space model within a modified U-Net. ExpoMamba is designed to address mixed-exposure challenges by decoupling the modeling of amplitude (intensity) and phase (structure) in the frequency domain. This allows for targeted enhancement, making it highly effective for real-time applications, including downstream tasks like object detection and segmentation. Our experiments on six benchmark datasets show that ExpoMamba is up to 2-3x faster than competing models and achieves a 6.8\% PSNR improvement, establishing a new state-of-the-art in efficient, high-quality low-light enhancement. Source code: https://www.github.com/eashanadhikarla/ExpoMamba.

Figures (11)

• Figure 1: [Left: 400x600; Middle: 3840x2160] Logarithmic Scatter Plot of Inference Time vs. PSNR. Baselines that used ground-truth mean information to produce metrics were reproduced without such information for fairness, [Right:] Spider chart with comparison of ExpoMamba outputs with other SOTA light-weight models on perceptual realism.
• Figure 2: Overview of the ExpoMamba Architecture. The diagram illustrates the information flow through the ExpoMamba model. The architecture efficiently processes sRGB images by integrating convolutional layers, 2D-Mamba blocks, and deep supervision mechanisms to enhance image reconstruction, particularly in low-light conditions.
• Figure 3: The effectiveness of various HDR tone mapping layers inside the FSS block. $\text{CSRNet}^{+}$ with shrunken conditional blocks and dilated convolutions lessens overexposed artifacts.
• Figure 4: Qualitative comparison of ExpoMamba and baselines on the LOLv1 dataset. Results demonstrate structural fidelity and color balance under mixed lighting conditions.
• Figure 5: Perceptual evaluation results for ExpoMamba and baseline models (FECNet, I-DRBN, I-SID, and IAT) across key criteria. ExpoMamba demonstrates a balanced and robust performance, particularly excelling in brightness, reduced blur, and detail retention.