OSMamba: Omnidirectional Spectral Mamba with Dual-Domain Prior Generator for Exposure Correction

TL;DR

This work tackles exposure correction under extreme real-world conditions by integrating frequency-domain modeling with generative priors. It introduces Omnidirectional Spectral Mamba (OSMamba), featuring OS-Scan to capture long-range dependencies across amplitude and phase spectra, and a Dual-Domain Prior Generator (DDPG) that leverages diffusion priors to restore lost details. A two-stage training regime uses a Dual-Domain Prior Extractor (DDPE) to learn priors from well-exposed references, followed by GT-free finetuning via DDPG to enable inference without ground-truth images. Empirical results on multiple exposure datasets show state-of-the-art PSNR/SSIM gains and superior qualitative restoration of illumination, structure, and fine textures. The approach offers a principled pathway for robust exposure correction with potential applications in HDR reconstruction and downstream vision tasks.

Abstract

Exposure correction is a fundamental problem in computer vision and image processing. Recently, frequency domain-based methods have achieved impressive improvement, yet they still struggle with complex real-world scenarios under extreme exposure conditions. This is due to the local convolutional receptive fields failing to model long-range dependencies in the spectrum, and the non-generative learning paradigm being inadequate for retrieving lost details from severely degraded regions. In this paper, we propose Omnidirectional Spectral Mamba (OSMamba), a novel exposure correction network that incorporates the advantages of state space models and generative diffusion models to address these limitations. Specifically, OSMamba introduces an omnidirectional spectral scanning mechanism that adapts Mamba to the frequency domain to capture comprehensive long-range dependencies in both the amplitude and phase spectra of deep image features, hence enhancing illumination correction and structure recovery. Furthermore, we develop a dual-domain prior generator that learns from well-exposed images to generate a degradation-free diffusion prior containing correct information about severely under- and over-exposed regions for better detail restoration. Extensive experiments on multiple-exposure and mixed-exposure datasets demonstrate that the proposed OSMamba achieves state-of-the-art performance both quantitatively and qualitatively.

Figures (7)

• Figure 1: Comparison of OSMamba with previous methods.(a) Previous frequency-domain approaches (e.g., huang2022fecnet) process spectrum with convolutions, limited by local receptive fields. (b) Proposed OSMamba enjoys global receptive field by using omnidirectional spectral scanning. It also leverages a dual-domain prior generator to generate and inject diffusion prior into the network.
• Figure 2: Visual comparison of proposed OSMamba with state-of-the-arts. The green, blue, and red boxes highlight our method's advantages over FECNet huang2022fecnet and COTF li2024real in correction quality.
• Figure 3: Illustration of proposed OSMamba.(a) The overall architecture of OSMamba consists of multiple levels of Omnidirectional Spectral State Space Block (OS-SSB). (b) Each OS-SSB contains an Omnidirectional Spectral SSM (OS-SSM) and an Attention Module. The OS-SSM performs Omnidirectional Spectral Scanning (OS-Scan) on the amplitude and phase of deep features using Amplitude Mamba and Phase Mamba. (c) The compact teacher prior $\mathbf{Z}$ extracted by Dual-Domain Prior Extractor (DDPE) and the diffusion prior $\mathbf{\hat{Z}}_{0}$ generated by Dual-Domain Prior Generator (DDPG) are input to guide all OS-SSMs in Stages 1 (Pretraining) and 2 (Finetuning), respectively. During inference, $\mathbf{Z}$ is replaced by a randomly sampled Gaussian noise, and $\mathbf{\hat{Z}}_{0}$ is generated based on a condition $\mathbf{D}$ to guide detail restoration.
• Figure 4: Illustration of our developed Omnidirectional Spectral Scanning (OS-Scan) mechanism, which incorporates four scanning methods working in Fourier domain: row-wise, column-wise, positive diagonal, and negative diagonal scanning. The red arrows represent the associations between the element highlighted in the box and other elements in the amplitude spectrum.
• Figure 5: Visual comparison of our method against previous state-of-the-art approaches. The first row displays comparisons on the multi-exposure dataset MSEC, while the second row presents comparisons on the mixed-exposure dataset LCDP.