HF-CIM: High-Frequency Cross-View Interaction

Updated 6 July 2026

The paper's main contribution is demonstrating that confining stereo interaction to high-frequency wavelet branches avoids feature entanglement and yields measurable improvements, with PSNR gains of approximately 0.56 dB.
HF-CIM employs multi-scale wavelet decomposition and a PAM-style epipolar attention module to accurately align and transfer complementary high-frequency details between left and right views.
This design effectively separates low-frequency illumination adjustment from high-frequency detail restoration, leading to enhanced texture, denoising, and overall image quality under low-light conditions.

Searching arXiv for the main paper and a few closely related cross-view/frequency-aware papers to ground the article with current citations. arXiv search query: "Wavelet-based Decoupling Framework for low-light Stereo Image Enhancement" The High-Frequency Guided Cross-View Interaction Module (HF-CIM) is a stereo interaction component introduced in WDCI-Net, a wavelet-based decoupling framework for low-light stereo image enhancement. Its defining property is that cross-view interaction is restricted to the high-frequency wavelet branches, rather than being performed across the entire latent feature space. Within this design, low-frequency content is handled primarily as an illumination-adjustment problem, whereas high-frequency content is treated as the locus of cross-view detail recovery, texture restoration, and denoising. HF-CIM therefore operationalizes the paper’s broader claim that low-light stereo enhancement benefits from separating illumination restoration from collaborative detail recovery (Du et al., 16 Jul 2025).

1. Definition within the wavelet-decoupled stereo framework

HF-CIM appears inside WDCI-Net, whose central architectural move is to avoid encoding low-light degradation in a single entangled latent space. The network first applies a $3\times3$ convolution to a left-view low-light image $I_L \in \mathbb{R}^{H\times W\times 3}$ , producing a shallow feature embedding

$F_{L0}\in\mathbb{R}^{H\times W\times C}.$

It then applies a three-level DWT to this feature map. At each level $i\in\{1,2,3\}$ , the decomposition yields one low-frequency feature map $F_{Li}$ and three high-frequency feature maps $(V_{Li},H_{Li},D_{Li})$ , with resolutions $\frac{H}{2}\times\frac{W}{2}$ , $\frac{H}{4}\times\frac{W}{4}$ , and $\frac{H}{8}\times\frac{W}{8}$ across the three levels. The same weight-shared processing is applied to the right image $I_R$ (Du et al., 16 Jul 2025).

At the image level, the wavelet decomposition is written as

$I_L \in \mathbb{R}^{H\times W\times 3}$ 0

where $I_L \in \mathbb{R}^{H\times W\times 3}$ 1 is the low-frequency component and $I_L \in \mathbb{R}^{H\times W\times 3}$ 2 are high-frequency components in the horizontal, vertical, and diagonal directions. In the actual network, HF-CIM operates on the learned high-frequency feature triplets after feature-space DWT, not directly on raw-image coefficients (Du et al., 16 Jul 2025).

Branch	Processing module	Primary role
Low-frequency	IAM	Illumination and color adjustment
High-frequency, level $I_L \in \mathbb{R}^{H\times W\times 3}$ 3	HF-CIM $I_L \in \mathbb{R}^{H\times W\times 3}$ 4 DTEM	Cross-view detail exchange, then enhancement and denoising
Reconstruction	Inverse wavelet transforms	Progressive synthesis of enhanced left/right outputs

This placement is important: HF-CIM is not a general fusion block spanning all branches. It is a branch-specific stereo interaction module applied only to the high-frequency streams at all three wavelet scales.

2. Rationale for restricting cross-view interaction to high-frequency branches

The module is motivated by two claims made in the WDCI-Net design. First, wavelet transform permits independent processing of low-frequency and high-frequency information. Second, after multi-level wavelet decomposition, swapping only the low-frequency components between low-light and normal-light images largely determines the perceptual illumination of the reconstruction. This is used to support the proposition that illumination adjustment can be handled primarily through low-frequency content (Du et al., 16 Jul 2025).

From that premise, the authors argue that existing stereo low-light enhancement methods are suboptimal when they perform interaction in a shared latent feature space or across the entire feature space. The stated problems are feature entanglement, where illumination, texture, and noise are mixed together, and noise sensitivity and shortcut learning, since unrestricted interaction may propagate irrelevant illumination or noise patterns instead of useful fine structures. HF-CIM is therefore designed to operate “within high-frequency branches rather than across the entire feature space,” with the specific aim of “effectively extract[ing] valuable image details from the other view” (Du et al., 16 Jul 2025).

The resulting division of labor is explicit. Illumination adjustment is assigned largely to the single-view low-frequency branch, while cross-view benefit is reserved for high-frequency content, where stereo redundancy can help recover edges, textures, and fine structures. This suggests a modular restoration philosophy: low-frequency information is the carrier of illumination correction, whereas high-frequency information is the carrier of collaborative stereo detail completion.

This design axis is not universal across cross-view models. For example, AttenGeo uses an iterative Cross-view and Cross-attention Module (CVCAM) to exchange contextual information between views, but it does so without explicit frequency decomposition or high-frequency guidance (Zhu, 31 Oct 2025). Conversely, MFAF for cross-view geo-localization explicitly separates low-frequency structural cues and high-frequency edge details, but its interaction remains implicit through shared embedding learning rather than a dedicated high-frequency cross-view transfer block (Liu et al., 16 Sep 2025). HF-CIM is distinctive in combining explicit wavelet-domain decoupling with explicit stereo interaction confined to the high-frequency subspace.

3. Architectural placement and multi-scale operation

HF-CIM is applied after wavelet decomposition and before DTEM. For each wavelet level $I_L \in \mathbb{R}^{H\times W\times 3}$ 5, the left/right high-frequency triplets

$I_L \in \mathbb{R}^{H\times W\times 3}$ 6

are fed into HF-CIM, which outputs interacted high-frequency features

$I_L \in \mathbb{R}^{H\times W\times 3}$ 7

These are then passed to the Detail and Texture Enhancement Module (DTEM) for further enhancement and denoising. The low-frequency branch does not pass through HF-CIM (Du et al., 16 Jul 2025).

The resulting execution order is:

shallow convolution,
three-level wavelet decomposition,
low-frequency path $I_L \in \mathbb{R}^{H\times W\times 3}$ 8 IAM,
high-frequency path at each scale $I_L \in \mathbb{R}^{H\times W\times 3}$ 9 HF-CIM,
HF-CIM outputs $F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 0 DTEM,
inverse wavelet reconstruction.

Because HF-CIM is instantiated at all three DWT scales, the paper describes it as enabling multi-scale cross-view interaction of high-frequency information. That multi-scale property is central to the module’s intended behavior: stereo detail borrowing is not confined to one resolution, but is distributed over progressively coarser wavelet bands (Du et al., 16 Jul 2025).

A related architectural detail concerns the low-frequency path. The network does not rely only on DWT low-frequency features; it also creates downsampled image features $F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 1 for $F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 2 using PixelUnshuffle + $F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 3 convolution, concatenates them with the corresponding low-frequency DWT features, and fuses them by convolution before IAM. This enrichment occurs outside HF-CIM, and reinforces the broader separation between low-frequency illumination modeling and high-frequency stereo interaction (Du et al., 16 Jul 2025).

4. Internal mechanism: PAM-style epipolar attention on fused high-frequency features

HF-CIM is based on the Parallax Attention Mechanism (PAM). The justification is the standard stereo assumption that corresponding points largely lie along the epipolar line, so correspondence search can be performed row-wise. The module has two stages: generation of disparity attention maps, and transfer/fusion of high-frequency features across views (Du et al., 16 Jul 2025).

In the first stage, the three directional high-frequency features in each view are fused using Selective Kernel Feature Fusion (SKFF) rather than being processed independently. The paper gives two motivations: reduced computation and noise suppression, since high-frequency wavelet coefficients are noise-sensitive. After fusion, the resulting left/right high-frequency features are normalized and linearly projected: $F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 4

$F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 5

with shapes stated as

$F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 6

The attention maps are then computed as

$F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 7

$F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 8

where $F_{L0}\in\mathbb{R}^{H\times W\times C}.$ 9 denotes matrix multiplication. The paper states that correlation is computed “between arbitrary positions along the epipolar line,” but does not specify the softmax dimension, the exact reshaping beyond the displayed tensor forms, or whether disparity-window masking is used inside HF-CIM (Du et al., 16 Jul 2025).

In the second stage, the same attention map—estimated from fused high-frequency features—is applied separately to each directional band. For right-view enhancement: $i\in\{1,2,3\}$ 0

$i\in\{1,2,3\}$ 1

$i\in\{1,2,3\}$ 2

and symmetrically for left-view enhancement: $i\in\{1,2,3\}$ 3

$i\in\{1,2,3\}$ 4

$i\in\{1,2,3\}$ 5

where $i\in\{1,2,3\}$ 6 denotes element-wise summation.

The resulting pattern is precise: correspondence is estimated from fused high-frequency information, the derived attention is then used to align and transfer each directional subband from the opposite view, and the transferred feature is added back to the original same-band feature as a residual fusion. The strict characterization given by the paper is therefore a PAM-style epipolar attention module operating on fused high-frequency features, followed by residual per-band feature transfer (Du et al., 16 Jul 2025).

5. Functional role, relation to DTEM, and empirical evidence

HF-CIM and DTEM are explicitly sequential and complementary: $i\in\{1,2,3\}$ 7 Their roles are differentiated. HF-CIM performs cross-view interaction, meaning stereo information exchange and detail borrowing from the other view. DTEM performs intra-view enhancement/refinement of the interacted high-frequency features, with emphasis on boosting detail and texture, suppressing remaining noise, and leveraging correlations among the vertical, horizontal, and diagonal subbands (Du et al., 16 Jul 2025).

HF-CIM does not have a dedicated loss term. It is trained indirectly through the overall WDCI-Net supervision: $i\in\{1,2,3\}$ 8 None of these losses explicitly supervise correspondence or attention inside HF-CIM (Du et al., 16 Jul 2025).

The most direct evidence for the module comes from the ablation table. With the full model, the reported results are:

Left: 26.790 / 0.834
Right: 26.823 / 0.834

Without HF-CIM:

Left: 26.231 / 0.792
Right: 26.251 / 0.792

This corresponds to drops of $i\in\{1,2,3\}$ 9 left PSNR, $F_{Li}$ 0 right PSNR, and $F_{Li}$ 1 SSIM on both sides. The authors interpret this as evidence that without HF-CIM the network “cannot capture features from the other view” (Du et al., 16 Jul 2025).

The implementation details relevant to HF-CIM are partial. Explicitly stated are: 3 wavelet levels, 3 high-frequency branches per level, interaction at all three scales, SKFF before attention, normalization + linear projection, and a PAM-style attention core. Also stated are the training platform and schedule: PyTorch, RTX 3090 GPU, Adam, batch size 20, training crop $F_{Li}$ 2, validation crop $F_{Li}$ 3, initial learning rate 0.0002, halved every 250 epochs, total 1000 epochs. Not stated are the channel dimension $F_{Li}$ 4, attention head count, exact normalization type, disparity range value, SKFF hyperparameters, or FLOPs attributable specifically to HF-CIM (Du et al., 16 Jul 2025).

6. Ambiguities, limitations, and relation to adjacent module designs

Several caveats are explicit. The paper includes a caption naming inconsistency: Figure 1(a) refers to “the first stage of the HF-SMM module,” but the surrounding text clearly describes HF-CIM. The paper also does not provide a controlled comparison against an alternative “full-space PAM” baseline within the same architecture, so the claim that high-frequency-only interaction is better than whole-space interaction is argued conceptually and supported indirectly through overall performance rather than by an isolated head-to-head test (Du et al., 16 Jul 2025).

HF-CIM-specific failure cases are not analyzed in depth. There are no explicit examples of correspondence failure, occlusion breakdown, or reflective/non-Lambertian failure. The authors state that the method as a whole still has limitations in multi-scale processing efficiency and extreme low-light conditions, and future work is directed toward more efficient wavelet variants, adaptive decoupling strategies, and better collaboration between low-frequency illumination and high-frequency texture. A plausible implication is that the present fixed decoupling strategy, including HF-CIM, may not yet be fully optimal in very challenging scenes (Du et al., 16 Jul 2025).

Within the broader literature represented by recent arXiv work, HF-CIM occupies a specific position. AttenGeo provides a strong template for cross-view interaction through iterative cross-attention, but without explicit high-frequency decomposition (Zhu, 31 Oct 2025). Wavelet-based Global-Local Interaction Network with Cross-Attention for Multi-View Diabetic Retinopathy Detection separates high-frequency lesion enhancement and cross-view fusion into distinct modules rather than one unified block (Hu et al., 25 Mar 2025). DAGNet combines a Frequency Domain Interaction Module with a later cross-view enhancement stage, but its published FDIM equations do not show explicit inter-view spectral coupling (Hong et al., 3 Feb 2025). In heterogeneous fusion, FMRFusion pairs learned base/detail decomposition with cross-view complementary interaction, but does not explicitly formulate the attention as high-frequency-gated (Zhoua et al., 6 Jun 2026). These comparisons suggest that HF-CIM’s most distinctive contribution is not merely the presence of frequency-aware processing or cross-view attention, but the decision to make high-frequency wavelet branches themselves the exclusive substrate for stereo interaction.

A second contrast concerns how strongly high-frequency information should be trusted. In sparse-view 3D Gaussian splatting, DWTGS argues that high-frequency signals are often under-constrained and should be treated cautiously, with low-frequency supervision as the reliable substrate and high-frequency behavior controlled by sparsity rather than dense matching (Nguyen et al., 21 Jul 2025). In 3DGS super-resolution, ConFi-GS similarly separates detail demand from detail reliability and uses cross-view high-frequency consistency as a filtering criterion before injection (Li et al., 24 May 2026). These works do not invalidate HF-CIM, but they do indicate a broader methodological tension: cross-view high-frequency transfer is powerful, yet may require reliability control when supervision is weak or geometry is poorly constrained. That tension is not explicitly resolved inside WDCI-Net.

In technical summary, HF-CIM is the stereo-detail exchange mechanism of WDCI-Net. After wavelet decomposition has separated illumination-dominated low-frequency information from detail-dominated high-frequency information, it estimates stereo correspondence from a fused high-frequency representation and uses the resulting parallax attention to transfer vertical, horizontal, and diagonal wavelet details from one view to the other via residual fusion. Its central design claim is that, for low-light stereo enhancement, cross-view interaction is most beneficial in the high-frequency space, while low-frequency illumination restoration can be handled independently (Du et al., 16 Jul 2025).