Haar Wavelet Convolution (HWConv) Overview

• HWConv is a technique that decomposes input feature maps into LL, LH, HL, and HH subbands using the 2-D Haar wavelet transform to capture both frequency energy and spatial details.
• It employs recursive decomposition on the low-frequency (LL) branch and progressive inverse reconstruction, enabling deep frequency-spatial coupling for improved target detection.
• Empirical studies demonstrate that HWConv enhances performance metrics such as IoU and precision in tasks like infrared small target detection compared to conventional pooling and convolution methods.

Searching arXiv for papers on Haar Wavelet Convolution and related wavelet-convolution modules. Haar Wavelet Convolution (HWConv) denotes a wavelet-aware CNN operator in which a 2-D Haar wavelet transform decomposes a feature map into $LL$, $LH$, $HL$, and $HH$ subbands so that frequency-domain energy and spatial details can be processed jointly before inverse reconstruction or downstream fusion. In the available literature, the term appears explicitly in "SWAN: Synergistic Wavelet-Attention Network for Infrared Small Target Detection," where HWConv is the encoder-side feature extractor and is described as a nested Haar-transform pipeline tailored to infrared small target detection; closely related but non-identical references are the wavelet frequency enhancement downsampling (WFED) block in ARFC-WAHNet and the Haar-wavelet-domain Wavelet Compressed Convolution (WCC) operator for $1\times1$ convolutions (Jing et al., 2 Aug 2025, Cui et al., 15 May 2025, Finder et al., 2022).

1. Terminological scope and family resemblance

Within the cited literature, HWConv is not a single canonical operator. The most direct use of the term is in SWAN, which introduces “Haar Wavelet Convolution (HWConv)” for “a deep, cross-domain fusion of the frequency energy and spatial details of small target” (Jing et al., 2 Aug 2025). ARFC-WAHNet does not introduce a module literally named HWConv, but its WFED module is described as “a fairly direct design reference for an HWConv-like wavelet downsampling/reconstruction block” because it uses fixed Haar decomposition, subband-specific enhancement, inverse reconstruction, and residual pooling fusion (Cui et al., 15 May 2025). WCC is also not named HWConv; it is a Haar-wavelet compressed realization of $1\times1$ convolution, with transform-domain coefficient selection and inverse reconstruction (Finder et al., 2022).

Paper	Module term	Relation to HWConv
SWAN	HWConv	Direct use of the term
ARFC-WAHNet	WFED	HWConv-adjacent downsampling block
Wavelet Feature Maps Compression for Image-to-Image CNNs	WCC	Haar-wavelet-domain $1\times1$ convolution

This distribution of terminology is important because “Haar Wavelet Convolution” can refer to at least three distinct design patterns. In SWAN it is a nested $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$ feature extractor. In ARFC-WAHNet it is better understood as a structured downsampling-and-enhancement module rather than a plain learned convolution layer. In WCC it is a compressed-domain point-wise operator rather than a general spatial convolution algebra.

SWAN also positions HWConv against earlier wavelet-CNN integrations such as WaveCNet and WA-CNN, characterizing those earlier integrations as shallow combinations based on parallel branches, concatenation, or attention weighting, whereas its own claim is “deep frequency-spatial coupling” through recursive decomposition of the low-frequency branch and inverse-transform-based progressive fusion (Jing et al., 2 Aug 2025).

2. Canonical formulation in SWAN

In SWAN, HWConv is introduced to address what the paper identifies as a core weakness of conventional convolution in IRSTD: ordinary convolutions “primarily capture local spatial patterns” and “struggle to distinguish the unique frequency-domain characteristics of small targets from intricate background clutter” (Jing et al., 2 Aug 2025). The paper further states that CNNs prefer high-frequency features, while “infrared imaging small target characteristics tend to favor low-frequency features,” making the $LL$ branch central to the design.

Architecturally, HWConv is the feature extraction module at the encoder side or early backbone side of SWAN. The network has three major components—HWConv, SSA, and RDCA—and “the input image undergoes a two-level nested fusion with HWConv.” The HWConv outputs form the feature maps $F_H$ that are then passed to SSA: $LH$0 and the SSA-enhanced features are subsequently fused by RDCA (Jing et al., 2 Aug 2025).

The HWConv mechanism is described as a staged nested transform. The input tensor is denoted by $LH$1, and the first decomposition is

$LH$2

Only the low-frequency branch is recursively decomposed: $LH$3 At each level, the four subbands are processed by a $LH$4 convolution: $LH$5 The paper then describes inverse wavelet reconstruction and progressive additive fusion from deeper to shallower levels. Because the published formula is malformed, the safest summary is the one stated directly in the source: HWConv performs Haar transform first, then $LH$6 convolution on the resulting subbands, then inverse wavelet transform and hierarchical fusion (Jing et al., 2 Aug 2025).

The two defining architectural ideas are therefore recursive decomposition of $LH$7 alone and progressive reconstruction with addition. The first emphasizes low-frequency semantic or energy structure; the second returns nested information to the current stage scale. This makes HWConv neither a plain convolution with Haar kernels nor a mere wavelet pooling layer. It is a nested DWT–Conv–IWT block whose stated purpose is to fuse “frequency-domain energy” and “spatial-domain local details.”

3. HWConv-adjacent downsampling in ARFC-WAHNet

ARFC-WAHNet provides a closely related but explicitly different design. The paper states that it does not introduce a module literally named HWConv, and the relevant component is WFED rather than a generic Haar convolution layer (Cui et al., 15 May 2025). WFED is inserted in the encoder and is described as replacing traditional max pooling: “This module utilizes 2-D Haar wavelet transform (2D-Haar-WT) and frequency enhancement downsampling to replace the traditional max pooling operation…”

The WFED pipeline begins with a convolutional preprocessing stage: $LH$8 where $LH$9 is a convolutional operation and $HL$0 is 2D-Haar-WT. The transform “sequentially applies wavelet-domain low-pass filter (LPF) and high-pass filter (HPF) along the horizontal and vertical directions, followed by downsampling by a factor of 2 in each dimension,” yielding $HL$1, $HL$2, $HL$3, and $HL$4 subbands. According to the authors’ stated shape semantics, each subband has “one-fourth the original number of channels” and “half the spatial resolution” (Cui et al., 15 May 2025).

WFED then separates the processing of high- and low-frequency content. The three detail subbands $HL$5, $HL$6, and $HL$7 are processed jointly by a Laplacian-style high-pass filter in the frequency domain, followed by SE and PA attention: $HL$8 The low-frequency branch $HL$9 is processed by a Gaussian low-pass filter, SE, and sigmoid suppression: $HH$0 The enhanced subbands are then reconstructed by inverse Haar transform: $HH$1 Finally, the reconstructed feature map is fused with a residual max-pooled shortcut: $HH$2

The architectural significance of WFED is that it replaces downsampling by a structured analysis–synthesis mechanism. The paper repeatedly frames it as a method to preserve information during hierarchical reduction, separate low-frequency background structure from high-frequency detail, enhance target-sensitive components, suppress clutter, and reconstruct a stronger downsampled representation. This is HWConv-adjacent, but it is broader than a plain Haar convolution layer because it includes explicit frequency-domain filtering, attention, inverse reconstruction, and residual pooled fusion (Cui et al., 15 May 2025).

4. Compressed-domain Haar convolution in WCC

WCC supplies a different interpretation of Haar-wavelet-based convolution. The paper does not name its method HWConv; it introduces “Wavelet Compressed Convolution (WCC)” as a method for high-resolution activation-map compression integrated with point-wise convolutions (Finder et al., 2022). Conceptually, however, it is close to a Haar-wavelet-domain convolution layer because it applies a Haar transform, compresses coefficients, performs learned $HH$3 channel mixing in the Haar domain, and reconstructs with the inverse transform.

The Haar transform in WCC is a fixed, efficient, orthonormal spatial transform applied separately to each activation-map channel. For a single channel $HH$4, the one-level transform is defined as a 2-D convolution with stride $HH$5 using four $HH$6 kernels corresponding to the standard $HH$7, $HH$8, $HH$9, and $1\times1$0 subbands: $1\times1$1 The inverse transform is implemented by transposed convolution with the same kernel bank: $1\times1$2 The method uses a multi-level Haar transform, and “in this work we use 3 levels in all the experiments.” As in SWAN, recursive decomposition is applied only to the low-pass branch (Finder et al., 2022).

The defining WCC operator is

$1\times1$3

Here $1\times1$4 is the Haar transform, $1\times1$5 is a top-$1\times1$6 shrinkage operator, $1\times1$7 is the learned point-wise convolution, $1\times1$8 reinserts retained coefficients into the full coefficient layout by zero-filling omitted entries, and $1\times1$9 is the inverse Haar transform. The key point is that WCC targets $1\times1$0 convolutions only: it does not define a general spatial convolution theorem in Haar space, does not learn wavelet filters, and does not restrict processing to $1\times1$1 alone. Instead, the retained coefficient vectors across all subbands are jointly processed after shared support selection across channels (Finder et al., 2022).

In this formulation, Haar wavelets are used primarily for activation compression and efficient point-wise channel mixing rather than for explicit target enhancement or attention. This marks a major divergence from IRSTD-oriented HWConv designs, even though the transform-domain skeleton—Haar analysis, subband-domain operation, inverse reconstruction—is closely related.

5. Empirical characterization

The strongest direct empirical evidence for HWConv comes from SWAN’s ablation studies. On IRSTD-Real, the baseline reports mIoU $1\times1$2, nIoU $1\times1$3, Pd $1\times1$4, Fa $1\times1$5, and F1 $1\times1$6; adding HWConv alone changes these to mIoU $1\times1$7, nIoU $1\times1$8, Pd $1\times1$9, Fa $1\times1$0, and F1 $1\times1$1. On NUDT, the baseline reports mIoU $1\times1$2, nIoU $1\times1$3, Pd $1\times1$4, Fa $1\times1$5, and F1 $1\times1$6; adding HWConv alone changes these to mIoU $1\times1$7, nIoU $1\times1$8, Pd $1\times1$9, Fa $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$0, and F1 $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$1. The same paper reports that two-level nesting is best overall, and that Haar gives the best overall tradeoff among Symlet, Coiflet, Biorthogonal, and Reverse Biorthogonal on the IRSTD-Real dataset (Jing et al., 2 Aug 2025).

Setting	Quantitative result	Context
Baseline $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$2 +HWConv on IRSTD-Real	mIoU $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$3, Pd $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$4, Fa $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$5	SWAN ablation
Baseline $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$6 +HWConv on NUDT	mIoU $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$7, nIoU $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$8, Fa $$\mathrm{Haar} \rightarrow \mathrm{Conv}_{3\times3} \rightarrow \mathrm{IWT}$$9	SWAN ablation
WFED vs alternatives	$LL$0 improves by $LL$1, $LL$2, $LL$3; $LL$4 increases by $LL$5, $LL$6, $LL$7	SIRST, NUDT-SIRST, IRSTD-1K

ARFC-WAHNet reports a dedicated WFED ablation comparing MaxPool, AvgPool, DWT, HWT, and WFED. The paper states that “MaxPool performs better than AvgPool for this task, but both are less effective than wavelet-based methods,” that “While both [Haar and Daubechies] yield similar performance, Haar is chosen for its simpler structure and higher computational efficiency,” and that “Among all methods, the proposed WFED achieves the best results on all three datasets.” Quantitatively, the reported gains are “$LL$8 improves by $LL$9, $F_H$0, and $F_H$1, while $F_H$2 increases by $F_H$3, $F_H$4, and $F_H$5 on SIRST, NUDT-SIRST, and IRSTD-1K, respectively” (Cui et al., 15 May 2025).

WCC provides a different empirical argument, centered on compression rather than IRSTD accuracy. On $F_H$6 activation maps from MobileNetV3-small on ImageNet, the paper reports much lower MSE for wavelet compression than for standard quantization at matched effective bit rate: at effective 2-bit, standard quantization MSE is $F_H$7 while wavelet compression MSE is $F_H$8; at effective 4-bit, the corresponding values are $F_H$9 and $LH$00. Across detection, segmentation, depth estimation, and super-resolution, the main quantitative message is that aggressive activation quantization causes sharp degradation, whereas WCC with 8-bit wavelet coefficients plus shrinkage degrades performance much more gracefully (Finder et al., 2022).

6. Misconceptions, implementation ambiguities, and limitations

A frequent misconception is to treat HWConv as a standard learned convolution kernel that merely happens to use Haar filters. The available papers do not support that simplification. In SWAN, HWConv is explicitly a Haar transform followed by $LH$01 convolution on subbands, followed by inverse wavelet reconstruction and fusion. In ARFC-WAHNet, the corresponding wavelet block is “not presented as a standard learned convolution kernel that simply happens to use Haar filters,” but as a “fixed wavelet decomposition + frequency-selective enhancement + inverse reconstruction + residual downsampling fusion module.” In WCC, the operator is a Haar-wavelet compressed implementation of $LH$02 convolution rather than a general-purpose wavelet convolution operator (Jing et al., 2 Aug 2025, Cui et al., 15 May 2025, Finder et al., 2022).

A second misconception is that Haar-wavelet modules necessarily emphasize only high-frequency detail. SWAN makes the opposite domain-specific argument for IRSTD, stating that CNNs prefer high-frequency features while infrared small target characteristics tend to favor low-frequency features, and it therefore recursively decomposes only the $LH$03 branch. ARFC-WAHNet uses both directions simultaneously: Laplacian-style enhancement and attention on $LH$04, $LH$05, and $LH$06, but Gaussian filtering, SE, and sigmoid suppression on $LH$07. This suggests that the role of low-frequency structure is task-dependent rather than secondary by definition (Jing et al., 2 Aug 2025, Cui et al., 15 May 2025).

The literature also leaves several implementation details unresolved. SWAN does not clearly specify subband packing, exact channel dimensions inside HWConv, stride, padding, normalization, or whether its $LH$08 convolutions are shared across subbands. ARFC-WAHNet does not provide explicit Haar kernel coefficients such as $LH$09 or $LH$10, and it does not state the padding strategy for HWT or IHWT. WCC is more explicit mathematically, but it still targets $LH$11 convolution only and is stated to be less beneficial when spatial resolution quickly shrinks to very small maps in standard image-classification pipelines (Jing et al., 2 Aug 2025, Cui et al., 15 May 2025, Finder et al., 2022).

Finally, the cited ablations indicate that more wavelet recursion is not always better. In SWAN, two-level nesting is optimal, while deeper nesting causes gradual degradation. The paper interprets this as a balance among feature abstraction, cross-modal interaction, and computational burden. A plausible implication is that HWConv behaves best when the recursion depth matches the scale at which low-frequency semantic stabilization remains useful without eroding local target evidence (Jing et al., 2 Aug 2025).