Wavelet-Space Feature Decomposition

• Wavelet-space feature decomposition is a multiscale representation that transforms data into the wavelet domain, distinctly separating low- and high-frequency components.
• The methodology employs fixed and adaptive transforms—such as DWT, SWT, and WTT—to effectively extract features in applications like EEG analysis, FTIR spectroscopy, and image classification.
• Practical implementations range from neural network subband regularization to operator compression, thus improving accuracy, efficiency, and robustness in various domain-specific tasks.

Searching arXiv for the cited papers to ground the article in current bibliographic records. {"query":"id:(Kharyuk et al., 2018) OR id:(Frikha, 2019) OR id:(Sinha et al., 2021) OR id:(Xiangyu, 2022) OR id:(Floryan et al., 2020) OR id:(Zhou et al., 17 Jan 2025) OR id:(Poudel et al., 22 Aug 2025)","max_results":10,"sort_by":"submittedDate"} I’ll also search for the exact topic phrase to ensure relevant wavelet-space decomposition literature is captured beyond the supplied IDs. {"query":"\"wavelet-space\" feature decomposition OR wavelet-based feature decomposition","max_results":10,"sort_by":"relevance"} Wavelet-space feature decomposition is a family of multiscale representations in which a signal, image, operator, or function is transferred to a wavelet domain and then processed through approximation/detail coefficients, critically sampled or shift-invariant subbands, localized basis functions, or hierarchical wavelet subspaces. Across the cited literature, the same general idea supports feature extraction for Fourier-transform infra-red spectra of medicinal plants, Stationary Wavelet Transform decomposition of EEG brainwave sub-bands, structurally regularized convolutional networks, sinogram-domain reconstruction, super-resolution, operator compression, and function-space decompositions on domains (Kharyuk et al., 2018, Frikha, 2019, Sinha et al., 2021).

1. Core representational forms

The most common construction is the Discrete Wavelet Transform. In two dimensions, the decomposition is expressed as

$$\mathbf{X}_{ll}=\mathbf{L}\mathbf{X}\mathbf{L}^{\mathrm T},\quad \mathbf{X}_{lh}=\mathbf{H}\mathbf{X}\mathbf{L}^{\mathrm T},\quad \mathbf{X}_{hl}=\mathbf{L}\mathbf{X}\mathbf{H}^{\mathrm T},\quad \mathbf{X}_{hh}=\mathbf{H}\mathbf{X}\mathbf{H}^{\mathrm T},$$

so that low-frequency and high-frequency components are explicitly separated in the feature map (Xiangyu, 2022). In image-classification architectures, the same operation is also written as a lossless, orthogonal decomposition into multiple critically sampled subbands,

$$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$

with each subband processed independently (Sinha et al., 2021).

The Stationary Wavelet Transform occupies a distinct place in this taxonomy. In EEG source localization it is described as preserving signal properties better than the Discrete Wavelet Transform due to the avoidance of downsampling (binary decimation), and in rendering super-resolution the stationary transform is used specifically to avoid spatial down-sampling, ensuring alignment across subbands and preserving shift invariance (Frikha, 2019, Poudel et al., 22 Aug 2025). By contrast, the Wavelet Tensor Train reinterprets decomposition in tensorized form. For FTIR spectra, WTT is compared directly with DWT and is distinguished by having only one parameter to be tuned, namely rank (Kharyuk et al., 2018).

Adaptive forms broaden the notion of wavelet-space decomposition beyond fixed filter banks. The Continuous Wavelet Transform provides a time-frequency representation with scale and translation parameters and is used for spectral fingerprinting and wavelet coherence analysis in audio identification (Shore, 1 Aug 2025). The empirical wavelet transform constructs adaptive wavelet filter banks from the averaged Fourier supports of the data, and the data-driven wavelet decomposition learns an orthogonal basis localized in scale and space, without imposing fixed basis functions or built-in self-similarity (Ren et al., 2020, Floryan et al., 2020).

Transform family	Representative form	Characteristic in the cited literature
DWT	$LL,LH,HL,HH$ or approximation/detail coefficients	critically sampled, orthogonal decomposition
SWT	full-resolution subbands	avoids downsampling; preserves shift invariance
WTT	tensor-train cores	one parameter to be tuned: rank
CWT / EWT / DDWD	scale-translation, empirical bands, learned hierarchy	adaptive time-frequency or data-driven multiscale analysis

2. Spectral and time-series feature extraction

In medicinal-plant spectroscopy, Fourier-transform infra-red spectra of samples from 7 plant species were used to explore the influence of preprocessing and feature extraction on efficiency of machine learning algorithms. Wavelet Tensor Train and Discrete Wavelet Transforms were compared as feature extraction techniques for FTIR data of medicinal plants; best results for WTT and DWT found through grid search were similar, significantly improving quality of clustering as well as classification accuracy for tuned logistic regression in comparison to original spectra, while WTT was presented as a more versatile and easier to use data processing tool because only rank must be tuned (Kharyuk et al., 2018).

In EEG, wavelet-space feature decomposition is embedded in a longer inverse-problem pipeline. A comparison of 51 mother wavelets from 7 different wavelet families was performed in a Stationary Wavelet transform decomposition of an EEG signal, extracting five different brainwave sub-bands for source localization. The workflow combines SWT, Independent Component Analysis for feature extraction, the Boundary Element Model for the forward problem, and the Equivalent Current Dipole for the inverse problem; the evaluation identified sym 20 as the best choice followed by bior 6.8 and coif 5 (Frikha, 2019).

Wavelet-space features also appear as nonlinear observables rather than as direct classifier inputs. In enhanced Dynamic-Mode Decomposition, wavelet based Besov norms are used to compute nontrivial multiscale nonlinear features of a given data set. The observables are defined from detail and approximation coefficients across scales and then appended to the canonical state variables for DMD, yielding a multiscale feature space that improves reconstruction accuracy and robustness to perturbations in the Nonlinear Schrödinger Equation setting (Curtis, 2020).

In time-frequency fingerprinting for audio identification, the Continuous Wavelet Transform is used to extract spectral features and wavelet coherence analysis compares recorded audio spectrograms to synthetically generated tunes derived from ABC notation. The method is reported to identify recorded tunes accurately and efficiently and is also discussed for EEG signal analysis and financial time series forecasting (Shore, 1 Aug 2025).

A deliberately unconventional example is parity detection. Integers are transformed into binary sequences, zero-padded, decomposed by the DWT up to level 3, and summarized by Energy, $L_2$ Norm, and Mean Absolute Value. K-means clustering with $k=2$ is then applied per feature, producing an approximately $69.67\%$ unsupervised accuracy. This suggests that wavelet-space feature decomposition can expose statistical regularities even in a purely discrete symbolic domain, while not recovering the modular rule itself (Mutlu, 29 Oct 2025).

3. Neural and statistical learning in wavelet space

A direct architectural use of wavelet-space feature decomposition appears in the Subband Regularized CNN. The proposed architecture decomposes the input image spectra into multiple critically sampled subbands, extracts features using a single CNN per subband, and performs classification by combining the extracted features using a fully connected layer. The architectural constraint that each CNN only processes a single subband is described as structural regularization. On ImageNet-2012, SRCNN (1 layer) achieves $65.6\%$ Top-1 and $82.17\%$ Top-5 accuracy, while requiring $169.5$M MACs, and the approach is also reported to be more robust than the regular full-band CNN to noise caused by weight-and-bias quantization and input quantization (Sinha et al., 2021).

WA-CNN introduces wavelet-space decomposition inside the network body rather than only at the input. Intermediate feature maps are decomposed into low-frequency and high-frequency components, where the low-frequency component stores the structures of the basic objects and the high-frequency components store detailed information and noise. The Wavelet-Attention block implements attention only in the high-frequency domain, using $\mathbf{X}_{lh}$ and $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$0 to generate an attention map that modulates $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$1. Based on MobileNetV2 backbones, WA-CNN achieves $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$2 Top-1 accuracy improvement on CIFAR-10 and $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$3 Top-1 accuracy improvement on CIFAR-100 (Xiangyu, 2022).

In sparse-view computed tomography, SWARM introduces a Sinogram-based Wavelet random decomposition And Random mask diffusion Model. The sinogram $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$4 is decomposed by DWT into $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$5, and a random training strategy is applied to the high-frequency components. The method couples this decomposition with a random mask strategy and a two-stage iterative reconstruction method, using one diffusion model for masked sinograms and another for randomly selected high-frequency subbands, with inverse wavelet recombination followed by FBP at inference (Zhou et al., 17 Jan 2025).

Wavelet-guided super-resolution extends the same principle to diffusion backbones. HDW-SR replaces conventional U-Net downsampling with wavelet-based downsampling, performs diffusion only on the residual map, and uses sparse cross-attention between the high-frequency subbands of the pre-super-resolved image and the low-frequency subbands of the diffused image. A Dynamic Thresholding Block refines high-frequency selection during sparse attention, and inverse wavelet transform is used during upsampling for low-loss feature reconstruction (Yang et al., 17 Nov 2025).

In real-time rendering, wavelet-space regression is used instead of RGB-space regression. A single-level 2D wavelet analysis operator decomposes the target image into $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$6, the network predicts these coefficients, and the inverse transform reconstructs the output. Ablations show that incorporating SWT improves PSNR by up to $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$7 dB and reduces LPIPS by $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$8 on average, with a computational overhead of roughly $$(X_0^{1},\ldots,X_0^{K})=\mathrm{DWT}(X_{\text{in}},K,M),$$9 ms compared to the DFASR baseline (Poudel et al., 22 Aug 2025).

4. Data-adaptive and domain-specific decompositions

Wavelet-space feature decomposition is not restricted to fixed analytic wavelet families. The data-driven wavelet decomposition discovers an energetic hierarchy of structures localized in scale and space by optimizing basis vectors under orthogonality and localization constraints. When applied to turbulence, the method reveals self-similar structure over a wide range of spatial scales, and the paper emphasizes that this self-similarity is not built into the basis but emerges from the data if present (Floryan et al., 2020).

The empirical wavelet transform provides another adaptive construction. In image-based flow decomposition, the two-dimensional tensor EWT decomposes instantaneous flow data or their visualizations according to averaged Fourier supports, producing modes that each account for part of the spectrum. The method is reported to separate jet and wake flows and their instabilities, and to robustly extract streamwise-elongated streaks, multiple secondary instabilities, and helical vortex filaments from both visualizations and full 3D flow data (Ren et al., 2020).

For 3D human faces, the wavelet transform is used as a statistical decorrelation and localization device. A second generation spherical wavelet transform decomposes the surface of the face, and many localized, decorrelated multilinear models are learned on the resulting coefficients. The resulting statistical shape space reconstructs faces from noisy and occluded 3D face scans and facial motion sequences, while preserving fine detail better than a global multilinear model and handling variation in expression better than a localized PCA model (Brunton et al., 2014).

Geometric Wavelet Gradient Boosting shifts the decomposition target from signals to tree predictors. Each child node $LL,LH,HL,HH$0 of a partition is assigned a geometric wavelet

$LL,LH,HL,HH$1

and any tree function is written as

$LL,LH,HL,HH$2

Wavelet norms order local contributions, and Out-Of-Bag data select the number of retained wavelets at each boosting stage. The reported effect is improved performance, particularly in scenarios of class unbalance and mislabeling in the training data (Dekel et al., 2018).

5. Operators, vector fields, and inverse problems

In operator compression, the object being decomposed is not a signal but a linear operator. Product-convolution expansions write

$LL,LH,HL,HH$3

and the wavelet representation is then

$LL,LH,HL,HH$4

The proposed algorithms compute wavelet decompositions of operators in quasi-linear time from a few impulse responses, which is positioned as a practical alternative to the $LL,LH,HL,HH$5 cost of naive wavelet decomposition of integral operators (Escande et al., 2020).

In weak gravitational lensing, wavelet-space feature decomposition is formulated through a Helmholtz split of the shear field,

$LL,LH,HL,HH$6

together with divergence-free and curl-free vector wavelets adapted to bounded domains. The method introduces border wavelets so that zero B-mode constraints can be imposed only at the boundary. In the reported experiments, the border-constrained wavelet Helmholtz method achieved relative RMS reconstruction errors as low as $LL,LH,HL,HH$7, compared to $LL,LH,HL,HH$8 for FFT and $LL,LH,HL,HH$9 for Seitz and Schneider, while reconstructing borders better than the existing methods considered there (Deriaz et al., 2012).

A plausible implication is that wavelet-space feature decomposition becomes especially attractive when the target object has geometric constraints—bounded domains, border artifacts, nonstationary local impulse responses, or coupled divergence/curl structure—that are awkward in purely Fourier-based formulations.

6. Functional-analytic foundations and conceptual scope

At the level of function spaces, wavelet-space feature decomposition becomes a structural theorem rather than a preprocessing step. In work following Triebel, reinforced spaces are introduced to treat critical values of the smoothness parameter $L_2$0 for Triebel-Lizorkin spaces on cellular domains, and the crucial decomposition theorems are transferred from non-critical to critical cases. The central decomposition separates an interior part from a boundary part through

$L_2$1

thereby enabling wavelet (Riesz) bases in reinforced function spaces where the classical construction fails (Scharf, 2013).

The Fourier-analytic reformulation of wavelet spaces extends this perspective. Feichtinger and Gröchenig coorbit spaces, defined through mixed weighted Lebesgue norms of the wavelet transform with respect to a quasi-regular representation of $L_2$2, are shown to be isomorphic under the Fourier transform to decomposition spaces $L_2$3. This equivalence provides a common framework for spaces associated to different dilation groups and makes geometric invariance questions transparent in the frequency covering $L_2$4 (Führ et al., 2014).

On product spaces of homogeneous type, Han, Li, and Ward’s product Hardy spaces are shown to admit an atomic decomposition built from the Auscher-Hytönen wavelet bases and their underlying reference dyadic grids. The consequences are that for all $L_2$5 the product atomic Hardy spaces coincide with the product Hardy spaces, and that these spaces are independent of the particular choices of both the wavelet bases and the reference dyadic grids (Han et al., 2018).

Representation-theoretic work extends the scope still further. General wavelet representations associated with finite-to-one endomorphisms on compact metric spaces admit a direct integral decomposition over a fundamental domain in the solenoid $L_2$6, and Martin boundary methods classify harmonic functions and commutants arising from the associated random walks and wavelet filters (Dutkay et al., 2011).

The surveyed literature therefore shows that wavelet-space feature decomposition is not synonymous with a single DWT preprocessing step. It includes critically sampled and stationary subband analysis, tensor-train and empirical constructions, learned hierarchical bases, decomposition of predictors and operators, vector-field splits, and function-space or representation-theoretic decompositions. A recurring technical theme is that high-frequency components can encode both detailed information and noise, while low-frequency components retain coarse structure; a recurring methodological theme is that the utility of the decomposition depends on how faithfully the wavelet domain matches the geometry, scale hierarchy, and invariances of the underlying problem.