Multi-scale Decomposition & Reconstruction

• Multi-scale decomposition reconstruction is a framework for separating data into scale-specific components and recombining them to enhance detail and regularization.
• It employs hierarchical, variational, and neural approaches to optimize both the decomposition and synthesis processes, ensuring accuracy and stability.
• Applications span image denoising, inverse problems, and multi-resolution analysis where scale alignment improves performance metrics like PSNR and SSIM.

Searching arXiv for papers on multi-scale decomposition reconstruction and closely related formulations. Multi-scale decomposition reconstruction denotes a family of procedures in which a signal, image, field, matrix, or latent representation is separated into constituents associated with different scales and then reassembled by summation, inverse transform, or composition. In the cited literature, this paradigm appears in hierarchical Tikhonov and total-variation decompositions, linear and learned image pyramids, block-wise low-rank matrix models, local multi-scale supervision for masked image modeling, coarse-to-fine diffusion priors, and multiresolution decompositions for inverse problems and dynamical systems (Wang, 2013, Canh et al., 2020, Ong et al., 2015, Wang et al., 2023, Thaker et al., 30 Jan 2026, Kutz et al., 2015). The common objective is not merely to partition data by resolution, but to align representation, regularization, and reconstruction with the scale structure of the underlying phenomenon.

1. Scale as a modeling variable

In the surveyed works, “scale” is defined operationally rather than abstractly. In learned image models, it may be the supervision patch size: LocalMIM defines scales $p_1<p_2<\dots<p_P$, partitions the input $x\in\mathbb R^{H\times W\times C}$ into non-overlapping $p_j\times p_j$ patches, and derives supervision $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$ from descriptors such as normalized pixels or HOG (Wang et al., 2023). In compressive imaging, it may be a filterbank output $x^{(s)}=D_s B_s x$, with Laplacian pyramid, Haar wavelet, or scale-space decompositions furnishing the scale channels (Canh et al., 2020). In Laplacian-pyramid diffusion models, the bands are $x^{(3)}=I^{(3)}$, $x^{(2)}=I^{(2)}-U(I^{(3)})$, and $x^{(1)}=I^{(1)}-U(I^{(2)})$ (Thaker et al., 30 Jan 2026).

Scale also appears as a partition of time, geometry, or matrix support. In mrDMD, level $\ell$ partitions the time axis into $J_\ell=2^{\ell-1}$ windows and extracts slow and fast components within each window (Kutz et al., 2015). In multi-scale low-rank matrix decomposition, each scale $x\in\mathbb R^{H\times W\times C}$0 is associated with a partition $x\in\mathbb R^{H\times W\times C}$1 of the matrix indices into blocks $x\in\mathbb R^{H\times W\times C}$2, and each component is locally low-rank on its own partition (Ong et al., 2015). In AMD for time series forecasting, repeated average-pooling generates scale patterns $x\in\mathbb R^{H\times W\times C}$3, and coarse scales are mixed back into fine ones in a residual top-down fashion (Hu et al., 2024).

This diversity shows that multi-scale decomposition reconstruction is not tied to a single representation theory. A plausible implication is that the notion of scale is most useful when it is coupled to the inductive bias of the task: patch size in masked prediction, block size in low-rank modeling, time bin in dynamical decomposition, or diffusion time in PDE-based filtering.

2. Reconstruction operators and synthesis rules

The reconstruction step is as central as the decomposition. In additive image decompositions, the original datum is recovered as a coarse residual plus detail bands. The SwV pipeline defines $x\in\mathbb R^{H\times W\times C}$4, $x\in\mathbb R^{H\times W\times C}$5, and reconstructs by

$x\in\mathbb R^{H\times W\times C}$6

where $x\in\mathbb R^{H\times W\times C}$7 controls detail amplification or suppression (Wong, 2021). The constrained-diffusion method similarly forms dyadic bands $x\in\mathbb R^{H\times W\times C}$8 and reconstructs

$x\in\mathbb R^{H\times W\times C}$9

with exact reconstruction up to finite-$p_j\times p_j$0 error (Li, 2022). In low-rank matrix decomposition, the recovered matrix is simply recombined as $p_j\times p_j$1 (Ong et al., 2015).

Other settings require non-additive synthesis. For manifold-valued data, Grohs and Wallner replace vector-space addition and subtraction by $p_j\times p_j$2, $p_j\times p_j$3, and geodesic weighted averages. Analysis computes

$p_j\times p_j$4

while synthesis reconstructs

$p_j\times p_j$5

The paper also shows a basic obstruction: for arbitrary nonlinear subdivision filters on manifolds, one cannot expect a general perfect-reconstruction theory analogous to the linear biorthogonal case; interpolating and midpoint-interpolating constructions are the cases that bypass this obstruction (Grohs et al., 2010). In image registration, the multiscale expansion is compositional rather than additive, with an optimal registration written as

$p_j\times p_j$6

instead of $p_j\times p_j$7 (Modin et al., 2018).

A common misconception is that multi-scale decomposition automatically implies exact invertibility. The literature is more specific: some frameworks provide exact reconstruction identities, some provide exact reconstruction up to numerical discretization error, and some admit only structured or local notions of invertibility.

3. Variational and hierarchical formulations

A large part of the theory begins with Tikhonov-type decompositions. Tadmor, Nezzar, and Vese’s hierarchical $p_j\times p_j$8 scheme, further analyzed in subsequent work, defines

$p_j\times p_j$9

with residuals $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$0. The paper proves monotonicity of the data fidelity, a telescoping energy identity, and $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$1-convergence $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$2 when $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$3 (Wang, 2013). In the base Tikhonov decomposition,

$y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$4

the component $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$5 captures the “good” part and $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$6 the residual (Wang, 2013).

The Banach-space MHDM generalizes this structure to linear ill-posed problems with fidelity

$y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$7

and regularizers $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$8 or $y^{(j)}\in\mathbb R^{H/p_j\times W/p_j\times D}$9. With $x^{(s)}=D_s B_s x$0, it iterates

$x^{(s)}=D_s B_s x$1

Under generalized triangle-inequality assumptions and geometric conditions on $x^{(s)}=D_s B_s x$2, the residual obeys a decay estimate and converges to zero; in the noisy case, stopping by a discrepancy principle yields stability (Kindermann et al., 2023). The same work shows that Bregman iteration can be written as an adaptive MHDM, and that coincidence with single-step Tikhonov regularization holds only under specific zero-Bregman-increment conditions; one-dimensional total-variation denoising is one such case (Kindermann et al., 2023).

The abstract multiscale theory of Modin, Nachman, and Rondi extends the paradigm to nonlinear inverse problems and diffeomorphic image registration. With increasing $x^{(s)}=D_s B_s x$3 and decreasing $x^{(s)}=D_s B_s x$4, the iterates

$x^{(s)}=D_s B_s x$5

yield convergence of the fidelities to the infimum value $x^{(s)}=D_s B_s x$6 and, under additional assumptions, convergence of the parameters (Modin et al., 2018).

Blind deconvolution introduces a coupled multiscale hierarchy for both image and kernel: $x^{(s)}=D_s B_s x$7 with scale-wise energies penalized by fractional Sobolev norms and a positivity constraint $x^{(s)}=D_s B_s x$8 to break the scale-indeterminacy $x^{(s)}=D_s B_s x$9. The method establishes residual convergence in the noise-free case and discrepancy-principle stability in the noisy case (Wolf et al., 2024).

4. Neural architectures for multi-scale decomposition and reconstruction

Recent learned systems internalize decomposition and reconstruction inside the architecture rather than treating them as separate preprocessing and inverse steps. LocalMIM is exemplary: it groups encoder layers into fine-scale layers $x^{(3)}=I^{(3)}$0 and coarse-scale layers $x^{(3)}=I^{(3)}$1, attaches a tiny decoder to each selected layer, and optimizes

$x^{(3)}=I^{(3)}$2

with default $x^{(3)}=I^{(3)}$3. Lower and upper layers reconstruct fine-scale and coarse-scale supervision signals respectively, and the default ViT-B layer set is $x^{(3)}=I^{(3)}$4 with HOG descriptors using 18 orientation bins (Wang et al., 2023).

MS-DCI jointly learns decomposition, sampling, and reconstruction. The decomposition stage is implemented by parallel convolutions $x^{(3)}=I^{(3)}$5, sampling by further convolutions $x^{(3)}=I^{(3)}$6, and reconstruction by an initial $x^{(3)}=I^{(3)}$7 pseudo-inverse-like layer followed by a five-layer enhancement module and a Multi-Level Wavelet CNN. The system is trained in three phases with losses $x^{(3)}=I^{(3)}$8, $x^{(3)}=I^{(3)}$9, and $x^{(2)}=I^{(2)}-U(I^{(3)})$0 (Canh et al., 2020).

MsDCNN addresses compressed sensing reconstruction with a fully-convolutional measurement operator and a Multi-scale Feature Extraction module. In the MFE, several parallel dilated-convolution branches with rates such as $x^{(2)}=I^{(2)}-U(I^{(3)})$1, $x^{(2)}=I^{(2)}-U(I^{(3)})$2, $x^{(2)}=I^{(2)}-U(I^{(3)})$3 extract multi-scale features from the same feature map, concatenate them, and fuse them by a $x^{(2)}=I^{(2)}-U(I^{(3)})$4 convolution (Wang et al., 2022).

Scale-cascaded diffusion models for medical super-resolution make the decomposition explicit. Three separate diffusion priors are trained: an unconditional prior for the coarsest band $x^{(2)}=I^{(2)}-U(I^{(3)})$5, a conditional prior for $x^{(2)}=I^{(2)}-U(I^{(3)})$6, and a conditional prior for $x^{(2)}=I^{(2)}-U(I^{(3)})$7. Posterior sampling proceeds from coarse to fine, using scale-specific data-consistency steps (Thaker et al., 30 Jan 2026).

Other domains adopt analogous ideas. SurfR precomputes per-cell features on parallel multi-scale grids with scales $x^{(2)}=I^{(2)}-U(I^{(3)})$8, fuses query-time features across scales by a Transformer encoder, and regresses sign logits and magnitude for the SDF (Ranade et al., 10 Jun 2025). AMD decomposes each time-series channel by repeated average-pooling, performs residual coarse-to-fine mixing, models temporal and channel dependencies, and uses adaptive multi-predictor synthesis to combine expert forecasts (Hu et al., 2024). VSRNN reconstructs unresolved physics as

$x^{(2)}=I^{(2)}-U(I^{(3)})$9

mirroring the variational multiscale expansion of the fine scales (Pradhan et al., 2021).

5. Empirical performance across domains

The empirical record is heterogeneous but substantial. In masked image modeling, LocalMIM reports that on ViT-B/16 with 100 pre-train epochs it reaches MAE’s $x^{(1)}=I^{(1)}-U(I^{(2)})$0 fine-tune accuracy in $x^{(1)}=I^{(1)}-U(I^{(2)})$1 GPU-hours versus MAE’s $x^{(1)}=I^{(1)}-U(I^{(2)})$2 GPU-hours. On Swin-B with 100 pre-train epochs, it reaches equivalent SimMIM accuracy $x^{(1)}=I^{(1)}-U(I^{(2)})$3 in 100 GPU-hours versus 360 GPU-hours. For ADE20K semantic segmentation with UperNet and ViT-B, MAE at 1600 epochs gives $x^{(1)}=I^{(1)}-U(I^{(2)})$4 mIoU, whereas LocalMIM-HOG at 1600 epochs and 1120 GPU-h gives $x^{(1)}=I^{(1)}-U(I^{(2)})$5 mIoU. For COCO detection and segmentation with Mask R-CNN and Swin-B, LocalMIM-HOG at 400 epochs gives $x^{(1)}=I^{(1)}-U(I^{(2)})$6 box AP and $x^{(1)}=I^{(1)}-U(I^{(2)})$7 mask AP (Wang et al., 2023).

In compressed sensing, MsDCNN-3 achieves mean PSNR $x^{(1)}=I^{(1)}-U(I^{(2)})$8 dB at MR=$x^{(1)}=I^{(1)}-U(I^{(2)})$9, $\ell$0 dB at MR=$\ell$1, and $\ell$2 dB at MR=$\ell$3, versus DR2-Net’s $\ell$4 dB and ReconNet’s $\ell$5 dB; SSIM is improved similarly by $\ell$6 (Wang et al., 2022). MS-DCI reports, on six $\ell$7 test images, that SS-DCI$\ell$8 reaches $\ell$9 at rate $J_\ell=2^{\ell-1}$0, $J_\ell=2^{\ell-1}$1 at rate $J_\ell=2^{\ell-1}$2, and $J_\ell=2^{\ell-1}$3 at rate $J_\ell=2^{\ell-1}$4, outperforming CSNet and S-CSNet in the reported comparisons (Canh et al., 2020).

For medical super-resolution, the three-level scale-cascaded diffusion model reports, on held-out FastMRI Brain, Knee, and Prostate slices, $J_\ell=2^{\ell-1}$5 PSNR, $J_\ell=2^{\ell-1}$6 SSIM, and $J_\ell=2^{\ell-1}$7 LPIPS for $J_\ell=2^{\ell-1}$8 SR with the 3-level cascade, compared with $J_\ell=2^{\ell-1}$9 for DiffPIR and $x\in\mathbb R^{H\times W\times C}$00 for Multi-Grid PnP; inference speed is improved by $x\in\mathbb R^{H\times W\times C}$01 because coarser scales use smaller images and smaller networks (Thaker et al., 30 Jan 2026).

In 3D full waveform inversion, TT-3DIFWI with M-SSIM reports GPU memory for INR training dropping from 98 GB to $x\in\mathbb R^{H\times W\times C}$02 MB on the synthetic Overthrust tests, and achieves the lowest absolute error maps and clearest stratigraphic interfaces among the compared methods (He et al., 22 Jun 2026). In surface reconstruction, SurfR reports average inference times on Thingi10K of $x\in\mathbb R^{H\times W\times C}$03 s at $x\in\mathbb R^{H\times W\times C}$04, $x\in\mathbb R^{H\times W\times C}$05 s at $x\in\mathbb R^{H\times W\times C}$06, and $x\in\mathbb R^{H\times W\times C}$07 s at $x\in\mathbb R^{H\times W\times C}$08; at $x\in\mathbb R^{H\times W\times C}$09, POCO takes $x\in\mathbb R^{H\times W\times C}$10 s and P2S $x\in\mathbb R^{H\times W\times C}$11 s, while average Chamfer is $x\in\mathbb R^{H\times W\times C}$12 for SurfR and $x\in\mathbb R^{H\times W\times C}$13 for POCO (Ranade et al., 10 Jun 2025).

In rotating turbulence inpainting, GPOD, EPOD, and GAN exhibit a nontrivial trade-off: EPOD and GAN achieve lower RMSE than GPOD for small and medium centered-square gaps, yet “the non-linear GAN does not outperform one of the linear POD techniques” in point-wise reconstruction, whereas the GAN better reproduces multi-scale statistics, non-Gaussian tails, and extreme events (Li et al., 2022).

6. Recurring limitations, artifacts, and methodological tensions

Several recurrent issues structure the field. First, band-limited decompositions may introduce artifacts at sharp transitions. In astronomical maps, wave transforms can produce negative ringing around sharp edges; the constrained-diffusion method was designed specifically so that, for $x\in\mathbb R^{H\times W\times C}$14, each band remains non-negative and artifacts around discontinuities are absent (Li, 2022). In detail enhancement, the SwV filter emphasizes a gradient-preserving property, claiming that no gradient-reversal or halo artifact can occur because the filter never overshoots local means (Wong, 2021).

Second, multiscale structure does not remove the need for careful regularization design. Multi-scale low-rank matrix decomposition notes that block partitions break translation invariance and can create block seams; cycle spinning is introduced to reduce these blocking artifacts (Ong et al., 2015). In blind deconvolution, the MHDM approach produces comparable results to a single-step variational method and a non-blind MHDM while requiring less laborious parameter tuning “at the price of more computations” (Wolf et al., 2024).

Third, theoretical equivalence across formulations is exceptional rather than generic. Grohs and Wallner show that a general manifold analogue of biorthogonal wavelets cannot be expected to possess perfect reconstruction, except for specific interpolating constructions (Grohs et al., 2010). The MHDM literature likewise shows that agreement with single-step Tikhonov regularization requires necessary and sufficient conditions and holds, for instance, in one-dimensional total variation denoising and certain $x\in\mathbb R^{H\times W\times C}$15-regularized settings (Kindermann et al., 2023).

Finally, nonlinear models do not dominate linear ones on every criterion. In turbulent-flow reconstruction, POD-based methods remain competitive in point-wise error, while GANs better preserve statistical multi-scale properties and extreme events (Li et al., 2022). This suggests that “reconstruction quality” in multiscale settings is intrinsically multi-criteria: fidelity to individual samples, faithfulness of scale-resolved statistics, stability of inverse solutions, computational burden, and the interpretability of the recovered components do not necessarily align.