Orthogonal Tunable Wavelet Units

• Orthogonal tunable wavelet units are learnable, parameterized filter-bank modules that ensure strict orthogonality and perfect reconstruction through structured lattice designs.
• They decompose feature maps into low- and high-frequency components via two-channel analysis, effectively replacing traditional pooling and stride-two convolution in CNN architectures.
• Empirical studies demonstrate improved performance in OCT classification and 3D segmentation, validating these units as robust alternatives for downsampling in vision applications.

Orthogonal tunable wavelet units are parameterized wavelet or filter-bank modules whose coefficients are adjustable while orthogonality, and frequently perfect reconstruction, are preserved by construction rather than by post hoc projection. In recent CNN literature, they appear as learnable replacements for stride-two convolution, downsampling, and pooling, implemented through two-channel analysis filter banks that retain both approximation and detail information; in broader wavelet theory, closely related objects are realized through Nyquist-derived orthogonal multiresolution analyses, unitary circuit factorizations, conservative linear systems, and structured autoencoder parameterizations (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025, Oliveira et al., 2015, Evenbly et al., 2016, Charina et al., 2016, Jawali et al., 2021).

1. Conceptual definition and distinguishing features

In the recent vision literature, a tunable wavelet unit is a learnable, wavelet-like module integrated into a CNN that uses a two-channel analysis filter bank, low-pass $H_0$ and high-pass $H_1$, to perform a 2D discrete-wavelet-transform-style decomposition of feature maps. It is tunable because the filter coefficients are learned during training, and it is orthogonal when the parameterization guarantees an orthogonal two-channel filter bank with perfect reconstruction for every admissible parameter value. In that role, the unit replaces or augments downsampling, stride-two convolutions, and pooling layers (Le et al., 1 Jul 2025).

The defining distinction from max-pooling, average pooling, or unconstrained strided convolution is structural. Max-pooling and average pooling reduce spatial resolution while discarding high-frequency content; unconstrained strided convolution learns a downsampling operator but does not, in general, preserve energy or provide a perfect-reconstruction interpretation. Orthogonal tunable wavelet units instead separate low-frequency structure and high-frequency detail explicitly, typically producing $LL$, $LH$, $HL$, and $HH$ subbands and then learning how to fuse them. In the OCT papers, this is presented as preserving “both low-pass and high-pass features”; in broader filter-bank terms, it is an orthogonal critically sampled analysis stage whose synthesis stage is either explicit or implicitly available by construction (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025).

A broader reading across the cited work suggests a common invariant: orthogonality is not treated as an empirical regularity but as a parameterization principle. Whether the parameters are lattice rotation angles, roll-off factors, circuit angles, Hermitian projections, or filter taps constrained by architectural ties, the target object is an orthogonal wavelet/filter-bank unit whose tunability does not destroy exact algebraic structure (Oliveira et al., 2015, Evenbly et al., 2016, Charina et al., 2016, Cotronei et al., 2012).

2. Orthogonal lattice parameterization in learned wavelet units

The most explicit recent orthogonal construction is OrthLatt-UwU, or OrthLattUwU, introduced as an orthogonal tunable wavelet unit in CNNs for OCT analysis and 3D retinal segmentation. Its analysis filters are built through a cascade of orthogonal rotation matrices and delay matrices, i.e. an orthogonal lattice structure. The trainable quantities are angles $\{\theta_k\}$, with each stage defined by

$$R_k= \begin{bmatrix} \cos(\theta_k) & \sin(\theta_k)\ -\sin(\theta_k) & \cos(\theta_k) \end{bmatrix}, \qquad \Lambda(z)= \begin{bmatrix} 1 & 0\ 0 & z^{-1} \end{bmatrix}.$$

Because each $R_k$ is orthogonal and the lattice form is inherited from orthogonal filter-bank theory, any admissible parameter vector yields an orthogonal filter bank; no explicit orthogonality regularizer is required (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025).

In the 3D segmentation formulation, orthogonality also fixes the remaining synthesis and analysis filters once the low-pass analysis filter $h_0$ is known: $H_1$0 This gives exact alias cancellation and perfect reconstruction in the orthogonal case, as well as energy preservation in the usual unitary sense up to downsampling (Le et al., 22 Jul 2025).

The actual feature-map transform follows classical separable wavelet analysis. If $H_1$1 and $H_1$2 denote low-pass and high-pass filtering-plus-downsampling matrices, then a feature map $H_1$3 is decomposed as

$H_1$4

This is the standard approximation/detail splitting of a 2D DWT. In the 2D ResNet18 setting, the four subbands are combined through a one-layer FCN,

$H_1$5

whereas in the 3D MGU-Net setting they are fused by an attention head. The underlying point is the same: downsampling is performed only after structured frequency separation (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025).

This orthogonal lattice design resolves a common misunderstanding that learnability and strict orthogonality are incompatible. In these constructions, tunability is confined to differentiable structural parameters—chiefly the lattice angles—so the model remains trainable by ordinary backpropagation while exact orthogonality is preserved for all parameter values (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025).

3. Integration into convolutional architectures

In the ERM surgery classification study, the baseline architecture is ResNet18. Traditional stride-two convolutions are replaced by a stride-one convolution followed by a tunable wavelet unit, and pooling layers are likewise replaced by tunable wavelet units. The paper states that these units replaced “traditional stride-convolution, downsampling, and pooling functions in ResNet18.” Filter lengths of 2, 4, 6, and 8 taps were tested, with the best reported main results obtained for 8-tap models (Le et al., 1 Jul 2025).

The 3D retinal-layer segmentation study adopts motion-corrected MGU-Net and replaces every max-pool layer in the encoder path with a wavelet unit. OrthLattUwU performs a 2D in-plane orthogonal DWT per channel and per depth slice, yielding $H_1$6, $H_1$7, $H_1$8, and $H_1$9 subbands of shape $LL$0, followed by attention-based fusion. The paper emphasizes that orthogonality implies simple analysis/synthesis relations, energy preservation, and exact perfect reconstruction, making the unit a theoretically grounded alternative to pooling (Le et al., 22 Jul 2025).

A related refinement is the stop-band-energy-constrained orthogonal unit introduced for texture-rich classification and anomaly detection. Here the underlying orthogonal lattice parameterization is retained, but the low-pass filter is additionally regularized by the normalized stop-band energy

$LL$1

and the total classification objective becomes

$LL$2

This adds a frequency-selective bias without sacrificing structural orthogonality, and is motivated by the observation that orthogonality alone does not force a learned low-pass channel to maintain clear stop-band attenuation (Le et al., 21 Jul 2025).

4. Empirical behavior, competing variants, and practical trade-offs

In postoperative OCT classification of ERM-only removal versus ERM+ILM removal, baseline ResNet18 reached $LL$3 mean test accuracy and a best single test run of $LL$4. Preprocessing with Energy-Crop $LL$5 and Wavelet Denoising with DB2 increased mean test accuracy to $LL$6 and the best single test run to $LL$7. Adding OrthLatt-UwU-8Tap raised mean test accuracy to $LL$8 and the best single test result to $LL$9, while PR-Relax-UwU-8Tap reached $LH$0 mean test accuracy and a best single test result of $LH$1. A trained ophthalmologist graded the same test set at $LH$2 accuracy (Le et al., 1 Jul 2025).

These results establish an important nuance. Orthogonality is useful, but it is not uniformly dominant over every relaxed alternative. PR-Relax-UwU uses exact alias cancellation through the quadrature mirror relation

$LH$3

but enforces perfect reconstruction only approximately through a half-band penalty added to cross-entropy,

$LH$4

In this OCT study, that additional flexibility produced slightly better accuracy than the strictly orthogonal model (Le et al., 1 Jul 2025).

The same pattern appears in 3D retinal segmentation. Baseline MGUNet with max-pooling achieved average Dice $LH$5 and average Accuracy $LH$6. OrthLattUwU variants were comparable or slightly better, with the best orthogonal configuration being the 2-tap Haar-initialized model at average Dice $LH$7 and average Accuracy $LH$8. However, the best overall result came from LS-BiorthLattUwU with 2 lifting steps, at average Dice $LH$9 and average Accuracy $HL$0 (Le et al., 22 Jul 2025).

Stop-band-energy-constrained orthogonal units show a different advantage profile. On CIFAR-10, baseline ResNet18 achieved $HL$1, while the best SBE-OrthLatt-UwU result was $HL$2 with DB2 and $HL$3. On DTD, baseline ResNet18 achieved $HL$4, whereas the best SBE-OrthLatt-UwU reached $HL$5 with Haar and $HL$6. In MVTec AD hazelnut, the SBE-OrthLatt-UwU encoder achieved segmentation AUROC $HL$7 and detection AUROC $HL$8, with the latter being the best among the compared encoders (Le et al., 21 Jul 2025).

A recurrent practical trade-off is therefore explicit in the literature: orthogonal units offer energy-preserving, stable, interpretable downsampling, but task-optimal performance may still favor PR-relaxed or biorthogonal variants, and stronger spectral regularization may be especially helpful on texture-dominated data (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025, Le et al., 21 Jul 2025).

5. Broader parameterization paradigms outside CNN-specific lattice units

Orthogonal tunable wavelet units are not confined to CNN pooling replacements. In “A Family of Wavelets and a new Orthogonal Multiresolution Analysis Based on the Nyquist Criterion” (Oliveira et al., 2015), orthogonality is obtained by taking the square root of a Nyquist raised-cosine spectrum, $HL$9, so that the Meyer partition-of-unity condition holds automatically. The roll-off factor $HH$0 becomes the tuning parameter, with the construction valid for $HH$1. This yields a family of orthogonal multiresolution analyses whose transition width and time-frequency localization are continuously tunable through $HH$2 (Oliveira et al., 2015).

A different viewpoint appears in the unitary-circuit literature. “Representation and design of wavelets using unitary circuits” (Evenbly et al., 2016) represents compact orthogonal dyadic transforms as products of local unitary gates $HH$3, while “Improved Wavelets for Image Compression from Unitary Circuits” (McCord et al., 2022) uses $HH$4 orthogonal reflection-symmetric gates $HH$5 in dilation-3 circuits. In both cases, orthogonality is exact because the transform is a product of local orthogonal blocks, and the tunable parameters are the gate angles. The cited papers use those parameters to recover Daubechies wavelets, symlets, coiflets, symmetric wavelets, multi-wavelets, and ternary wavelets optimized for image compression (Evenbly et al., 2016, McCord et al., 2022).

System theory provides a still more abstract description. “System theory and orthogonal multi-wavelets” (Charina et al., 2016) shows that compactly supported orthogonal wavelet and multi-wavelet masks can be identified with the transfer function

$HH$6

of a conservative linear system, where the block matrix built from $HH$7 is unitary. This recasts orthogonal wavelet design as the parameterization of unitary colligations. Closely related Lie-group parameterizations appear in “Partial parameterization of orthogonal wavelet matrix filters” (Cotronei et al., 2012), where FIR matrix wavelet filters are generated from orthogonal operators commuting with shifts by two and from $HH$8 parameters via the Lie algebra $HH$9 (Charina et al., 2016, Cotronei et al., 2012).

Learning-based wavelet design reaches a similar destination from another direction. “Learning optimal wavelet bases using a neural network approach” (Søgaard, 2017) treats the DWT as a structured linear network whose free parameters are low-pass filter coefficients, with orthonormality enforced by quadratic regularization terms corresponding to the usual wavelet conditions. “Wavelet Design in a Learning Framework” (Jawali et al., 2021) instead uses linear convolutional autoencoders and architectural parameter tying to learn orthogonal or biorthogonal filter banks from Gaussian data, with vanishing moments imposed through B-spline factorization and symmetry handled either as a hard architectural constraint or as a soft penalty (Søgaard, 2017, Jawali et al., 2021).

6. Misconceptions, limitations, and current research directions

One common misconception is that orthogonality alone determines the best learned wavelet unit. The comparative results do not support that claim. PR-Relax-UwU slightly exceeds OrthLatt-UwU on the ERM surgery task, and biorthogonal variants outperform the orthogonal variant in 3D retinal segmentation; strict orthogonality is therefore best understood as a structural prior rather than as a universal optimum criterion (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025).

A second misconception is that exact symmetry, compact support, orthogonality, and arbitrarily many vanishing moments can always be achieved simultaneously. The learning framework in (Jawali et al., 2021) reports the classical obstruction in practical form: for orthogonal, compactly supported wavelets with more than one vanishing moment, exact symmetry is infeasible except for the Haar case. That is why Symmlets are treated as approximately symmetric orthogonal wavelets, while exact symmetry is recovered more naturally in biorthogonal constructions such as CDF 9/7 (Jawali et al., 2021).

Current limitations are also explicit. In the OCT classification study, the dataset is small—340 training/validation samples and 50 test samples—and the paper identifies OCT volumes, data augmentation, and diffusion techniques as future directions. In 3D segmentation, gains over max-pooling are modest and longer orthogonal filters do not necessarily help; the 8-tap orthogonal model was slightly worse than shorter alternatives. In the stop-band-energy work, inference time rises from $\{\theta_k\}$0 s per image for baseline ResNet18 to $\{\theta_k\}$1 s for SBE-OrthLatt-UwU, indicating nontrivial overhead (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025, Le et al., 21 Jul 2025).

The broader literature points toward two complementary trajectories. One is architectural generalization: the retinal and texture papers present orthogonal tunable wavelet units as drop-in downsampling modules for 2D and 3D CNNs, and the stop-band-energy paper states that future extensions are planned for Transformer architectures (Le et al., 22 Jul 2025, Le et al., 21 Jul 2025). The other is richer parameterization: Nyquist-based roll-off control, unitary-circuit factorization, conservative-system transfer functions, and autoencoder-based filter learning all show that orthogonality can be preserved exactly while leaving meaningful continuous degrees of freedom for optimization (Oliveira et al., 2015, Evenbly et al., 2016, Charina et al., 2016, Jawali et al., 2021).

In that sense, orthogonal tunable wavelet units are less a single layer type than a design principle: constrain the analysis/synthesis operator so that orthogonality and perfect reconstruction are structural, then optimize only within that admissible manifold. The recent CNN instantiations make that principle operational for downsampling in vision systems, while the older wavelet-theoretic constructions provide the larger mathematical vocabulary in which those units are situated (Le et al., 1 Jul 2025, Le et al., 22 Jul 2025, Evenbly et al., 2016, Charina et al., 2016).