---
title: Orthogonal Matrix Image Transformation (OMIT)
url: https://www.emergentmind.com/topics/orthogonal-matrix-image-transformation-omit
type: topic
---

# Orthogonal Matrix Image Transformation (OMIT)

Orthogonal Matrix Image Transformation (OMIT) is a synthesized designation for image and tensor mappings implemented by orthogonal matrices, together with the associated orthonormal basis-image or transform-domain factorization. In its classical 2-D form, an image \(F\) is mapped as \(G = U^\top F V\) and inverted as \(F = U G V^\top\); in multidimensional settings, a tensor \(\mathcal{X}\) can be transformed along a chosen mode by an orthogonal matrix \(Q\), as in \(\mathcal{T}_Q(\mathcal{X}) = \mathcal{X} \times_3 Q\). Across the literature, this framework appears in several guises: basis-image decompositions and block operators for digital arrays [1901.07840], constructive synthesis of discrete orthogonal transform matrices from symmetric sample sets [2101.05438], orthogonal low-complexity DCT approximations for image compression [1402.6034], and learnable orthogonal transforms for low-rank tensor inverse problems [2412.11165].

## 1. Algebraic definition and orthogonality

The defining property of OMIT is orthogonality. A real square matrix \(U\) or \(Q\) is orthogonal if \(U U^\top = I\) or \(Q^\top Q = I\). In 1-D, the corresponding transform pair is
\[
g = U^\top f, \qquad f = U g.
\]
In 2-D, for an image \(F \in \mathbb{R}^{M \times N}\) and orthogonal matrices \(U \in \mathbb{R}^{M \times M}\), \(V \in \mathbb{R}^{N \times N}\),
\[
G = U^\top F V, \qquad F = U G V^\top.
\]
In the square separable case emphasized in the basis-image literature, \(M=N\) and \(U=V\), yielding \(G = U^\top F U\) and \(F = U G U^\top\) [1901.07840].

Orthogonality implies exact invertibility and norm preservation. For vectors, \(\|f\|_2^2 = \|g\|_2^2\); for images, the analogous preserved quantity is the Frobenius norm. This is the algebraic basis for interpreting OMIT as an energy-preserving image transformation rather than merely a change of coordinates [1901.07840].

The same structure extends naturally to tensors. For a 3-way tensor \(\mathcal{X} \in \mathbb{R}^{n_1 \times n_2 \times n_3}\), the learned orthogonal-transform form used in low-rank inverse problems applies
\[
\mathcal{T}_Q(\mathcal{X}) = \mathcal{X} \times_3 Q, \qquad
\mathcal{T}_Q^{-1}(\hat{\mathcal{X}}) = \hat{\mathcal{X}} \times_3 Q^\top.
\]
Because \(Q\) is orthogonal, inversion is immediate through \(Q^\top\), and the transform remains linear, differentiable, and invertible [2412.11165].

## 2. Basis images, decomposition, and data scattering

A central classical interpretation of OMIT is the basis-image decomposition. Let \(u_k\) denote column \(k\) of \(U\). The basis images are defined by the outer products
\[
a_{kp} = u_k \otimes u_p, \qquad a_{kp}(x,y) = U_{xk} U_{yp}.
\]
Then the image admits the orthonormal expansion
\[
F = \sum_{k,p} a_{kp} G_{kp}, \qquad G_{kp} = \langle F, a_{kp} \rangle,
\]
with Frobenius inner product \(\langle A,B\rangle = \sum_{x,y} A_{xy} B_{xy}\). In the rectangular case with distinct \(U\) and \(V\), the separable basis images become \(B_{kp} = u_k \otimes v_p\) and the same decomposition principle applies [1901.07840].

These basis images are orthonormal:
\[
\langle a_{kp}, a_{mn} \rangle = \delta_{km}\delta_{pn}.
\]
Consequently,
\[
\sum_{k,p} G_{kp}^2 = \sum_{x,y} F_{xy}^2,
\]
which is the Parseval-type energy identity for the basis-image domain. The completeness relations
\[
\sum_k a_{kk} = I, \qquad \sum_x a_{kp}(x,x) = \delta_{kp}
\]
show that the set \(\{a_{kp}\}\) spans the full matrix space [1901.07840].

The paper on basis images describes orthogonal transforms as performing data scattering, meaning pixel energy redistribution across transform coefficients. In 1-D, each input component contributes to all transform coefficients through the entries of the orthogonal matrix. In 2-D, this redistribution occurs over the coefficient matrix \(G\) or, equivalently, over the basis-image coefficients \(\{G_{kp}\}\). Energy is preserved globally, but its distribution depends on the chosen transform. The same source notes that DCT, WHT, and KLT tend to compact energy for natural images, whereas DST does not [1901.07840].

A further extension uses a block matrix \(b\) whose entries are basis images, enabling orthogonal transforms of block vectors whose components are themselves matrices. In that block formalism,
\[
g = b f, \qquad f = b g,
\]
and the resulting transformed components can become non-separable even when the original block components are separable. The extraction identity
\[
\langle a_{kp}, g_p \rangle = \psi_k
\]
formalizes a correlation-producing property of the transform that was used to motivate a frequency-domain watermarking detection scheme recoverable from a single spatial block projection [1901.07840].

## 3. Constructing orthogonal transform matrices

OMIT is not restricted to pre-existing transforms such as DCT or WHT. One line of work gives a constructive method for generating an \(N \times N\) orthogonal matrix \(Q\) directly from a symmetric discrete sample set, without discretizing continuous orthogonal functions. For \(N=2m\), one selects \(m\) distinct positive values \(y_0,\dots,y_{m-1}\) and uses the symmetric grid \(\{-y_0,\dots,-y_{m-1},+y_0,\dots,+y_{m-1}\}\). Two polynomial families are then constructed: even polynomials \(P_{2t}(x)\) containing only even powers and odd polynomials \(P_{2t+1}(x)\) containing only odd powers. Symmetry ensures that cross-parity terms cancel, while same-parity terms are enforced to be orthogonal on the half-grid. The coefficients are obtained iteratively from linear systems \(A_t D_t = -B_t\) and \(\tilde A_t \tilde D_t = -\tilde B_t\), and the resulting polynomials are normalized so that \(2 \sum_{k=0}^{m-1} P_i(y_k)^2 = 1\) [2101.05438].

The construction has a proof of existence and uniqueness by induction, with \(\det(A_t)\neq 0\) and \(\det(\tilde A_t)\neq 0\) for all required degrees. The assembled matrix \(Q\) is formed by evaluating the normalized polynomials on the symmetric grid, using parity to fill the negative and positive halves efficiently. In this framework, the DCT-II arises by choosing
\[
y_k = \cos\!\left(\frac{(2k+1)\pi}{2N}\right),
\]
while the DTT arises from an evenly spaced arithmetic sequence on \((0,1)\); the same method also generates valid custom transforms from triangular numbers, primes, or Fibonacci values [2101.05438].

A different construction targets low-complexity compression rather than generic synthesis. The orthogonal 8-point DCT approximation defines an integer sign/zero matrix
\[
C_0 = \operatorname{round}(2C)
\]
from the exact DCT matrix \(C\), then orthogonalizes it via
\[
A = S C_0,
\]
where \(S\) is diagonal and chosen so that \(A A^\top = I\). The rows of \(C_0\) are mutually orthogonal with squared row norms \([8,6,4,6,8,6,4,6]\), and the runtime transform core uses only additions and subtractions because the irrational diagonal scaling is merged into quantization [1402.6034].

| Construction family | Core mechanism | Noted properties |
|---|---|---|
| Discrete polynomial synthesis | Even/odd discrete-orthogonal polynomials on \(\pm y_k\) | Generates DCT-II, DTT, and custom orthogonal transforms |
| 8-point DCT approximation | \(A = S C_0\) with \(C_0=\operatorname{round}(2C)\) | 22 additions, 0 multiplications, 0 shifts per 1-D transform |
| Learnable orthogonal transform | Product of Householder reflections | Exact orthogonality with data-adaptive learning |

The constructive synthesis method has offline complexity approximately \(O(N^4)\), whereas applying the resulting matrix in a separable 2-D transform is \(O(N^2)\) per block. By contrast, the 8-point DCT approximation was evaluated on 45 grayscale \(512 \times 512\) images from the USC-SIPI dataset, and its summed integrated spectral error relative to the exact DCT was reported as 1.79, compared with 3.32 for SDCT, 5.93 for BAS-2008, and 26.40 for BAS-2011; the transform also outperformed SDCT at all compression ratios in the reported experiments and outperformed BAS-2008 at both low-\(r\) and high-\(r\) operating points [2101.05438] [1402.6034].

## 4. Learned OMIT in multidimensional inverse problems

Recent work extends OMIT from fixed image transforms to learnable orthogonal transforms inside tensor low-rank models. In this setting, the orthogonal matrix \(Q \in \mathbb{R}^{n \times n}\) is parameterized as a product of Householder reflections,
\[
Q = \prod_{i=1}^k H(\mathbf v_i), \qquad
H(\mathbf v) = I - \frac{2\,\mathbf v \mathbf v^\top}{\|\mathbf v\|_2^2}.
\]
Each \(H(\mathbf v)\) is symmetric and orthogonal, so the product is exactly orthogonal for any values of the learnable vectors \(\{\mathbf v_i\}\). In practice, a normalized vector \(\tilde{\mathbf v} = \mathbf v / (\|\mathbf v\|_2+\epsilon)\) is used for stability [2412.11165].

This parameterization is used to define a transform-induced tensor algebra. For tensors \(\mathcal A \in \mathbb{R}^{n_1 \times \ell \times n_3}\) and \(\mathcal B \in \mathbb{R}^{\ell \times n_2 \times n_3}\), the transformed t-product is
\[
\mathcal A * \mathcal B
:= \mathcal T_Q^{-1}\!\big(\mathcal T_Q(\mathcal A)\triangle \mathcal T_Q(\mathcal B)\big),
\]
where \(\triangle\) denotes frontal-slice matrix multiplication in the transformed domain. The corresponding transform-induced t-SVD has the form
\[
\mathcal X = \mathcal U * \mathcal S * \mathcal V^\top,
\]
with a standard matrix SVD performed independently on each frontal slice of \(\hat{\mathcal X}=\mathcal T_Q(\mathcal X)\) [2412.11165].

The associated tensor nuclear norm is
\[
\|\mathcal X\|_{\mathrm{TNN},Q}
:= \sum_{i=1}^{n_3}
\left\|
\big(\mathcal T_Q(\mathcal X)\big)^{(i)}
\right\|_*,
\]
and yields the inverse-problem formulation
\[
\min_{\mathcal X}\;
\frac12\|\mathcal A(\mathcal X)-\mathbf y\|_2^2
+\lambda \|\mathcal X\|_{\mathrm{TNN},Q}.
\]
The stated motivation is that learning \(Q\) from data adapts the transform to the actual spectral or temporal correlation structure and can produce stronger energy compaction and lower effective tubal rank than fixed transforms such as DFT, DCT, or DWT [2412.11165].

Within this framework, OMIT becomes an explicitly task-aligned transform layer for multispectral images, multi-frame videos, and related inverse problems. A plausible implication is that the classical basis-image intuition of orthogonal scattering is being reinterpreted in modern tensor form as adaptive low-rank compaction in a learned transformed domain.

## 5. Solvers, optimization, and implementation

Two solution paradigms are described for learned tensor OMIT. The first is the classical convex-proximal route based on transformed-domain singular value soft-thresholding. For a proximal-gradient step with \(f(\mathcal X)=\tfrac12\|\mathcal A(\mathcal X)-\mathbf y\|_2^2\) and step size \(\eta>0\),
\[
\mathcal X^{t+1}
=
\mathrm{prox}_{\eta\lambda \|\cdot\|_{\mathrm{TNN},Q}}
\Big(\mathcal X^t - \eta \nabla f(\mathcal X^t)\Big).
\]
Operationally, this means transforming a gradient step into the \(Q\)-domain, performing slice-wise SVDs, shrinking singular values by \(\eta\lambda\), and inverting with \(Q^\top\) [2412.11165].

The second paradigm is the paper’s generative t-SVD parameterization, designed to avoid differentiating through SVD altogether. The unknown low-rank tensor is represented through factors \(\mathcal U \in \mathbb R^{n_1\times r\times n_3}\), \(\mathcal V \in \mathbb R^{n_2\times r\times n_3}\), a rank matrix \(S \in \mathbb R^{n_3\times r}\), and a dense rank estimation operator \(\rho(\cdot)\) implemented as an MLP. A simple form given in the source is
\[
\rho(\mathbf S)=\mathrm{LReLU}(\mathbf S W_1)W_2.
\]
The transformed-domain reconstruction is then
\[
\hat{\mathcal U}=\mathcal T_Q(\mathcal U), \qquad
\hat{\mathcal V}=\mathcal T_Q(\mathcal V), \qquad
\hat{\mathcal S}=\mathrm{Diag}(\rho(\mathbf S)),
\]
\[
\hat{\mathcal X}
=
\hat{\mathcal U}
\triangle
\hat{\mathcal S}
\triangle
\hat{\mathcal V}^\top,
\qquad
\mathcal X = \mathcal T_Q^{-1}(\hat{\mathcal X}).
\]
This realizes a learned, adaptive shrinkage in the transformed domain without backpropagating through eigenvectors or singular vectors [2412.11165].

The training objective combines task fidelity with an Orthogonal Total Variation regularizer:
\[
\mathrm{OTV}(\Theta)
=
\|\nabla_x \mathcal T_Q(\mathcal U)\|_1
+
\|\nabla_y \mathcal T_Q(\mathcal V)^\top\|_1
+
\|\nabla_x \theta_{Q^\top}\|_1.
\]
The fidelity term depends on the task: tensor completion uses \(\|\mathbf H_\Omega(\mathcal X)-\mathbf H_\Omega(\mathcal Y)\|_F^2\), CASSI reconstruction uses \(\|\mathbf X-\mathbf H(\mathcal Y)\|_F^2\), and denoising uses \(\|\mathcal X-\mathbf H(\mathcal Y)\|_1\) [2412.11165].

Implementation trade-offs are explicit. Materializing \(Q\) densely through \(k\) Householder multiplications costs \(O(k n^3)\), whereas applying stored reflection vectors implicitly costs \(O(k n^2)\) per fiber through the update \(y \leftarrow y - 2 v (v^\top y)/(v^\top v)\). The tensor transform \(\mathcal T_Q(\mathcal X)=\mathcal X\times_3 Q\) costs \(O(n_1 n_2 n_3 k)\) when reflections are applied implicitly and \(O(n_1 n_2 n_3^2)\) if a dense \(Q\) is precomputed. Slice-wise SVDs scale as
\[
O\!\left(\sum_{i=1}^{n_3} \min(n_1,n_2)^2 \max(n_1,n_2)\right),
\]
while skinny t-SVD with tubal rank \(r\) reduces this to \(O(n_1 n_2 n_3 r)\) [2412.11165].

## 6. Applications, empirical behavior, and limitations

The application space of OMIT spans both classical signal processing and contemporary inverse problems. In the basis-image formulation, the cited applications are image coding, watermarking, basis-wavelet synthesis, and feature extraction through interpretable transform coefficients [1901.07840]. In compression-oriented orthogonal approximations, OMIT appears as a block transform inside JPEG-like pipelines, with the forward 2-D transform \(Y = A X A^\top\), quantization scaling absorbed into the quantizer, and inverse reconstruction through the transpose [1402.6034]. In learned tensor settings, the applications are tensor completion, spectral imaging reconstruction with CASSI, and multispectral image denoising [2412.11165].

The reported datasets and metrics are specific. Tensor completion used CAVE multispectral images of size \(256\times256\times31\) and NTT videos consisting of the first 30 frames of \(288\times352\), evaluated by PSNR and SSIM. CASSI reconstruction used KAIST scenes of size \(256\times256\times31\) with shift \(d=2\), also evaluated by PSNR and SSIM. MSI denoising used KAIST scenes with Gaussian noise \(\sigma \in \{0.2,0.3\}\), evaluated by PSNR, SSIM, and FSIM [2412.11165].

The learned OMIT-based OTLRM model was reported to improve PSNR by about \(1\)–\(4\) dB over strong baselines including fixed-transform TNN, UTNN, DTNN, LS2T2NN, and HLRTF, especially at low sampling rates such as \(\mathrm{SR}=0.05 \rightarrow 0.15\). One example given is the Balloons scene at \(\mathrm{SR}=0.05\), where OTLRM achieved approximately \(40.20\) dB versus the next best approximately \(38.35\) dB. For CASSI, self-supervised OTLRM was reported to match or exceed supervised deep methods and to be on par or higher than the best supervised baselines in average PSNR over five KAIST scenes, with notably sharper boundaries. For MSI denoising, it was described as competitive with or better than model-based, Plug-and-Play, and diffusion-based methods, with qualitative improvements in texture fidelity and smoothness [2412.11165].

Several limitations recur across the literature. Energy compaction is transform-dependent: the basis-image source explicitly notes that DCT, WHT, and KLT tend to compact energy for natural images, whereas DST does not [1901.07840]. The discrete-polynomial construction requires even \(N=2m\), distinct positive sample values \(y_k\), and uniform weights \(w_k=1\) [2101.05438]. The learned low-rank tensor model assumes a strong transform-induced low-rank structure and may degrade when inter-band correlations are weak or structures are highly non-stationary; it also becomes computationally expensive when many reflections or dense per-slice SVDs are used, and its relaxed semi-orthogonality for \(\mathcal U,\mathcal V\) sacrifices some t-SVD optimality guarantees [2412.11165]. In the watermarking setting, robustness to severe geometric attacks or desynchronization is not addressed [1901.07840].

A common misconception is to treat OMIT as a single fixed transform comparable to the DCT. The cited material indicates a broader picture. In one reading, OMIT is the general orthogonal-array framework based on basis images and block operators; in another, it is a constructive family of discrete orthogonal matrices; in a more recent reading, it is a learnable orthogonal transform layer embedded in proximal or generative tensor solvers. The unifying invariant across these variants is not a particular matrix, but exact orthogonality, invertibility by transpose, and transform-domain structure that can be exploited for scattering, compaction, compression, denoising, reconstruction, or correlation-inducing block operations [1901.07840] [2101.05438] [1402.6034] [2412.11165].

Source: https://www.emergentmind.com/topics/orthogonal-matrix-image-transformation-omit