Wavelet-Based Low-Light Stereo Enhancement

• The paper demonstrates the efficacy of decoupling illumination and detail via multi-level wavelet transforms, leading to improved stereo image restoration.
• The methodology combines stereo redundancy and frequency-domain processing to adjust low-frequency brightness while refining high-frequency textures and suppressing noise.
• Empirical results show significant improvements in PSNR, SSIM, and NIQE across synthetic and real low-light datasets, validating the proposed framework.

Wavelet-based low-light stereo image enhancement denotes a class of methods that restore a low-light stereo pair by combining stereo redundancy with frequency decomposition. In explicit wavelet formulations, each view is decomposed as $\{cA,cH,cV,cD\}=DWT(I)$, where the low-frequency component carries most illumination information and the high-frequency components carry directional detail; enhancement then proceeds by adjusting illumination in low-frequency branches, refining texture and suppressing noise in high-frequency branches, and reconstructing the enhanced pair with inverse wavelet transforms. The problem is driven by the joint presence of very low brightness and contrast, strong color cast, heavy signal-dependent and signal-independent noise, and, in real data, additional blur, together with the stereo-specific difficulty that cross-view interaction can amplify noise when correspondence is estimated in a noisy feature space (Zhao et al., 2024, Du et al., 16 Jul 2025).

1. Problem formulation and degradation structure

Low-light stereo image enhancement is commonly posed as a supervised image-to-image regression problem on rectified stereo pairs. A representative formulation uses inputs $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$ and predicts enhanced outputs $E_l, E_r \in \mathbb{R}^{H \times W \times 3}$, with corresponding normal-light ground truth images $E_l^G, E_r^G$ (Hu et al., 2024). In this setting, enhancement implicitly targets brightness restoration, contrast improvement, noise suppression, structure and texture preservation, and stereo consistency (Hu et al., 2024).

The stereo setting differs materially from monocular low-light enhancement because each view contains complementary information about the same 3D scene. This cross-view redundancy can recover structures that are weak or corrupted in one view but cleaner in the other. At the same time, low-light noise disturbs feature encoding and cross-view matching: encoders may respond to noise rather than structure, and parallax-attention-style interaction may align and reinforce noise patterns instead of real features, causing noise amplification during decoding (Zhao et al., 2024). This makes the representation space in which stereo interaction is performed a central design decision.

A recurring misconception is that stereo enhancement gains arise simply from adding a second image. The recent literature instead treats stereo interaction as conditional on the reliability of the underlying feature space. In low-light stereo, correspondence is beneficial only when the representation has already suppressed a sufficient fraction of noise or separated illumination from detail, which is why low-frequency priors, wavelet decompositions, or explicit frequency-domain constraints recur across the field.

2. Wavelet-domain decoupling of illumination and texture

The defining operation in explicit wavelet-based stereo enhancement is the discrete wavelet transform

$\{cA, cH, cV, cD\} = DWT(I),$

where $cA$ is the approximation component and $cH, cV, cD$ are directional high-frequency components (Du et al., 16 Jul 2025). In WDCI-Net, this decomposition is applied recursively for three levels, yielding low-frequency maps $F_{Li}$ and directional high-frequency maps $V_{Li}, H_{Li}, D_{Li}$ at progressively coarser scales. The paper empirically verifies the decoupling claim by swapping low-frequency components between low-light and normal-light images after 3-level DWT and reconstructing via inverse DWT; the reported behavior is that illumination distribution is largely encoded in the low-frequency components (Du et al., 16 Jul 2025). On that basis, the framework decomposes the feature space into a low-frequency branch for illumination adjustment and multiple high-frequency branches for texture enhancement, and states that no prior work had performed explicit frequency-domain decoupling for stereo enhancement before this formulation (Du et al., 16 Jul 2025).

This low-/high-frequency division is reinforced by related monocular evidence. WalMaFa reports that swapping the wavelet low-frequency $LL$ component between low-light and normal-light images yields larger brightness improvement and higher SSIM than swapping Fourier amplitude, while swapping Fourier phase produces better preservation or addition of fine details than swapping wavelet high-frequency bands (Tan et al., 2024). The model therefore assigns global brightness enhancement to wavelet $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$0 and local texture refinement to Fourier phase (Tan et al., 2024). Wave-Mamba reaches a closely aligned conclusion from another direction: most content information of an image exists in the low-frequency component, less in the high-frequency component, and the high-frequency component exerts a minimal influence on the outcomes of low-light enhancement (Zou et al., 2024).

Taken together, these results establish the central rationale of wavelet-based stereo enhancement. Low-frequency branches are used to model illumination, global contrast, and large-scale color structure, whereas high-frequency branches concentrate edges, textures, and much of the noise. This separation is not merely architectural convenience; it is the mechanism by which stereo enhancement avoids encoding illumination, detail, and degradation into a single entangled latent space.

3. Cross-view interaction in wavelet and low-frequency spaces

WDCI-Net operationalizes stereo interaction only in the high-frequency branches. Each view first undergoes shallow $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$1 convolution to produce $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$2, followed by three-level DWT. The low-frequency branch is enriched by PixelUnshuffle-based downsampled RGB features $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$3, processed by the Illumination Adjustment Module, while each scale’s high-frequency triplet is sent to the High-Frequency Guided Cross-View Interaction Module (HF-CIM) and then to the Detail and Texture Enhancement Module (DTEM); multi-level inverse DWT reconstructs the enhanced stereo pair. Left and right branches share weights, but cross-view coupling is confined to the high-frequency subbands (Du et al., 16 Jul 2025).

HF-CIM first fuses directional high-frequency features with Selective Kernel Feature Fusion, then computes stereo attention maps: $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$4

$I_l, I_r \in \mathbb{R}^{H \times W \times 3}$5

and applies the resulting attention to each directional subband, for example

$I_l, I_r \in \mathbb{R}^{H \times W \times 3}$6

with analogous updates for $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$7 and $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$8 in both directions (Du et al., 16 Jul 2025). The stated motivation is that illumination varies across views and is better handled independently, whereas geometry, edges, and corners are concentrated in the high-frequency structure, so matching in HF space is less affected by illumination interference (Du et al., 16 Jul 2025).

A closely related stereo line does not use explicit DWT but arrives at a similar representation-first principle. LFENet introduces a low-frequency information enhanced module based on a side window filter, concatenates the filtered low-frequency image with the original image, reweights channels with squeeze-and-excitation, and performs cross-view and cross-scale interaction in the resulting low-frequency enhanced image space. The paper explicitly states that this perspective is directly relevant to wavelet-based designs because wavelets also separate low-frequency approximation and high-frequency detail components before processing (Zhao et al., 2024).

By contrast, SDI-Net is frequency-aware but not wavelet-based. It uses two identical U-Net-style encoder-decoder branches with a Cross-View Sufficient Interaction Module at $I_l, I_r \in \mathbb{R}^{H \times W \times 3}$9, and incorporates frequency-domain information only through an FFT-based loss; the paper explicitly states that no wavelet transform, DWT, or inverse DWT is used in the architecture itself (Hu et al., 2024). This distinction matters: frequency-domain supervision and wavelet-domain processing are related but not identical design choices.

4. Illumination adjustment, detail restoration, and denoising modules

Within the wavelet stereo setting, the low-frequency branch is treated as an illumination sub-task. In WDCI-Net, the Illumination Adjustment Module has two residual stages. The first uses large-kernel convolutions to capture long-range dependencies and global illumination distribution, with SimpleGate for adaptive feature scaling. The second applies channel attention, again with SimpleGate, to reweight channels according to their contribution to illumination restoration (Du et al., 16 Jul 2025). The output at level 3, $E_l, E_r \in \mathbb{R}^{H \times W \times 3}$0 and $E_l, E_r \in \mathbb{R}^{H \times W \times 3}$1, is explicitly supervised against the 3-level low-frequency components of the normal-light ground truth, reinforcing the interpretation of the low-frequency branch as an illumination-and-color pathway rather than a general restoration backbone (Du et al., 16 Jul 2025).

High-frequency refinement is handled separately. DTEM first applies depthwise separable convolution and SKFF to obtain a fused high-frequency descriptor $E_l, E_r \in \mathbb{R}^{H \times W \times 3}$2, then uses cross-attention to enhance vertical and horizontal components,

$E_l, E_r \in \mathbb{R}^{H \times W \times 3}$3

and refines the diagonal component by

$E_l, E_r \in \mathbb{R}^{H \times W \times 3}$4

before a final depthwise separable convolution produces $E_l, E_r \in \mathbb{R}^{H \times W \times 3}$5 (Du et al., 16 Jul 2025). This two-stage sequence—cross-view aggregation followed by intra-view directional refinement—treats denoising as a structured discrimination problem inside the high-frequency branches.

Monocular wavelet LLIE supplies several parallel module designs that are directly relevant to this decomposition. Wave-Mamba proposes a Low-Frequency State Space Block for global content restoration in low-frequency subbands and a High-Frequency Enhance Block that uses enhanced low-frequency information to correct high-frequency information (Zou et al., 2024). OWDiff, also monocular, attributes the blocking artifacts of DiffLL to the non-overlapped Haar Wavelet Transform and replaces it with an Overlapped WT built from biorthogonal wavelets, together with a low-frequency-guided HFEBlock for sharper edges and more reliable textures (Peng et al., 9 Jun 2026). These results do not constitute stereo methods, but they show that wavelet-based low-/high-frequency specialization can be implemented with CNN, SSM, or diffusion backbones and with different wavelet families.

5. Frequency-domain supervision, datasets, and empirical behavior

Wavelet-based stereo enhancement is usually trained with mixed spatial and frequency constraints. WDCI-Net optimizes

$E_l, E_r \in \mathbb{R}^{H \times W \times 3}$6

where $E_l, E_r \in \mathbb{R}^{H \times W \times 3}$7 is an FFT-domain $E_l, E_r \in \mathbb{R}^{H \times W \times 3}$8 loss on the final outputs, $E_l, E_r \in \mathbb{R}^{H \times W \times 3}$9 is an SSIM loss on the final outputs, and the three $E_l^G, E_r^G$0-resolution terms supervise the low-frequency branch with FFT, SSIM, and VGG19 perceptual constraints (Du et al., 16 Jul 2025). The training set contains 1,289 synthetic stereo pairs assembled from uniform-illumination and non-uniform-illumination subsets of Flickr1024, Holopix50k, and KITTI; validation uses 30 pairs; testing uses 312 Flickr2014 pairs, 266 KITTI2015 pairs, and 191 real Holopix50k pairs. The implementation uses PyTorch, NVIDIA RTX 3090, Adam, batch size 20 for training, $E_l^G, E_r^G$1 random crops, an initial learning rate of $E_l^G, E_r^G$2 halved every 250 epochs, and 1000 total epochs (Du et al., 16 Jul 2025).

On synthetic Flickr2014, WDCI-Net reports $E_l^G, E_r^G$3 PSNR/SSIM for the left view and $E_l^G, E_r^G$4 for the right; on KITTI2015 it reports $E_l^G, E_r^G$5 and $E_l^G, E_r^G$6; on real Holopix50k it reports NIQE $E_l^G, E_r^G$7, all described as best among the listed baselines (Du et al., 16 Jul 2025). Ablation results further attribute measurable degradation to removing DTEM, HF-CIM, IAM, or any of the loss terms, supporting the necessity of both wavelet decoupling and stereo HF interaction (Du et al., 16 Jul 2025).

Related stereo work converges on the same role for frequency supervision even when explicit wavelets are absent. LFENet trains with

$E_l^G, E_r^G$8

where $E_l^G, E_r^G$9 is an FFT-domain $\{cA, cH, cV, cD\} = DWT(I),$0 loss and $\{cA, cH, cV, cD\} = DWT(I),$1 is an SSIM loss, and reports that omitting either term degrades PSNR/SSIM (Zhao et al., 2024). SDI-Net similarly uses

$\{cA, cH, cV, cD\} = DWT(I),$2

with $\{cA, cH, cV, cD\} = DWT(I),$3 defined by FFT-domain $\{cA, cH, cV, cD\} = DWT(I),$4, and its ablation shows that removing FFT loss drops PSNR/SSIM dramatically below even a simplified interaction baseline (Hu et al., 2024). The common pattern is that stereo interaction alone is insufficient; photometric enhancement must be stabilized by explicit spectral constraints.

6. Broader frequency-domain context, limitations, and research directions

Wavelet-based low-light stereo enhancement sits within a broader frequency-domain LLIE lineage. R2-MWCNN replaces pooling and upsampling by DWT and inverse DWT in a U-Net-style architecture, uses multi-level discrete wavelet transform to divide feature maps into distinct frequencies, and combines pixel, global, edge, and channel-wise losses for simultaneous contrast enhancement and denoising (Chen et al., 2023). WalMaFa uses Wavelet-based Mamba Blocks in the encoder and decoder and Fast Fourier Adjustment Blocks in the latent space, explicitly assigning wavelet $\{cA, cH, cV, cD\} = DWT(I),$5 to global brightness enhancement and Fourier phase to local detail refinement (Tan et al., 2024). SDTL compresses features through wavelet transform before a diffusion transformer and introduces Structure Enhancement Module and Structure-guided Attention Block to emphasize texture-rich tokens while reducing interference from noisy areas (Yin et al., 21 Apr 2025). A zero-shot line moves diffusion to the wavelet low-frequency domain and combines wavelet and Fourier priors during reverse sampling, but it is single-image and includes no geometric modeling (He et al., 2024).

These adjacent methods clarify what wavelet-based stereo enhancement is not. It is not simply “frequency-aware” enhancement, because FFT loss alone does not constitute wavelet processing. It is not merely monocular wavelet LLIE applied independently to each view, because independent per-view enhancement can break stereo consistency. And it is not equivalent to full-latent-space stereo attention, because several stereo papers explicitly argue that low-light interaction should occur in a denoising-friendly or decoupled feature space rather than in a single entangled representation.

Current limitations are also explicit. WDCI-Net notes that multi-scale processing efficiency and enhancement under extreme low-light remain challenging, that performance may depend on the chosen wavelet, and that under strong misalignment or occlusion HF-CIM attention may be less accurate (Du et al., 16 Jul 2025). OWDiff shows, in the monocular diffusion setting, that non-overlapped Haar WT can cause visible block-edge discontinuities and that higher wavelet levels may reduce computation but lead to over-smoothed results when guidance becomes too weak at low resolution (Peng et al., 9 Jun 2026). A plausible implication is that future stereo systems will increasingly combine wavelet-domain decoupling, low-frequency-guided high-frequency refinement, and explicit cross-view consistency mechanisms, while also revisiting the choice of wavelet family, overlap strategy, and geometry-aware conditioning.

In that sense, wavelet-based low-light stereo image enhancement has evolved into a distinct design paradigm: illumination is treated as a low-frequency restoration problem, detail transfer is restricted to high-frequency structure where stereo cues are most informative, and frequency-domain supervision is used to prevent enhancement from becoming a purely black-box mapping. The resulting systems are both more interpretable than single-space latent models and more specialized to the combined photometric and geometric structure of low-light stereo data.