EquivAnIA: A Spectral Method for Rotation-Equivariant Anisotropic Image Analysis

Conference: CVPR 2026 arXiv: 2603.11294 Code: GitHub Area: Medical Imaging / Image Analysis Keywords: anisotropic analysis, rotation equivariance, cake wavelets, ridge filters, angular registration

TL;DR

This paper proposes EquivAnIA, which employs a family of oriented filters (cake wavelets and ridge filters) to estimate the angular distribution of an image via weighted averaging in the frequency domain, replacing conventional angular binning. The method achieves truly numerically rotation-robust anisotropic analysis, with a dominant orientation estimation error of only 0.03° on synthetic images and a CT registration error of only 0.02°.

Background & Motivation

Anisotropic analysis is ubiquitous in medical and scientific imaging — for determining tissue fiber orientation, detecting structural orientation in CT scans, and analyzing preferred directions in material textures. The core analytical tool is the 2D power spectral density (PSD), whose angular PSD \(S(\theta)\) is obtained by integrating along the radial direction in polar coordinates, encoding the power distribution across orientations.

Fatal flaw of conventional methods: On a discrete Cartesian grid, angular binning approximates \(S(\theta)\) by assigning each frequency point to a bin based on its angle and summing within bins. However, the Cartesian grid is inherently anisotropic — the bin at 0° contains far more frequency points than the bin at 30°. This leads to a critical consequence: the same image yields different angular distributions after rotation, i.e., the method lacks rotation equivariance. Registration errors can reach 20°, which is entirely unacceptable in medical imaging.

EquivAnIA's approach: Replace hard bin boundaries with a family of oriented filters defined as smooth functions in the frequency domain. The smooth weighted averaging of the filters eliminates quantization errors caused by the discrete grid. Combined with a radially symmetric windowing step to handle non-disk-supported images, the method achieves true rotation equivariance. The approach is purely spectral and contains no learnable parameters.

Method

Overall Architecture

A three-step pipeline: (1) PSD estimation — apply a radially symmetric window to non-disk-supported images before computing the FFT periodogram; (2) Directional filtering — perform weighted averaging of the PSD in the frequency domain using a filter bank; (3) Angular distribution extraction — obtain the energy response \(\rho(\theta)\) per orientation and identify the dominant direction via argmax.

Key Designs

1. Oriented Filter Bank Replacing Angular Binning

2. Function: Replace the hard partitioning of discrete bins with smooth frequency-domain filters that perform weighted averaging at each angular direction.

3. Mechanism: A filter bank \(\phi_{v,\theta}(u) = \phi(R_\theta^{-1}(u-v))\) is generated by rotating a base function \(\phi\). Analysis coefficients \(c_{v,\theta}\) are computed, and the angular distribution is defined as the energy per direction: \(\rho(\theta) = \int |c_{v,\theta}|^2 dv\). Two concrete filter types are used: Cake wavelets (sector-shaped coverage, suited for structural images) and Ridge filters (line-shaped, suited for texture images).

4. Design Motivation: The frequency-domain weights of the filters are continuous and smooth, eliminating quantization errors at bin boundaries. Rotating the input is equivalent to rotating the filter bank, so the analysis result is naturally rotation-equivariant.

5. Radially Symmetric Window Preprocessing

6. Function: Apply smooth windowing to rectangular images with non-disk support.

7. Mechanism: A radially symmetric window function with approximate disk support is applied, discarding information in image corners that may enter or exit the field during rotation.

8. Design Motivation: The corner regions of a rectangular image change during rotation, leading to inconsistent PSD estimation. Although discarding corner information sacrifices some data, it ensures stability in rotational analysis. Experiments show that periodogram estimation outperforms Bartlett/Welch methods — angular resolution is more critical than noise reduction.

9. Angular Image Registration Algorithm

10. Function: Estimate the relative rotation angle between two rotated copies of the same image.

11. Mechanism: The dominant orientations \(\hat{\theta}^{(1)}, \hat{\theta}^{(2)}\) of both images are computed independently. Since the method cannot distinguish \(\theta\) from \(\theta + \pi\), two candidate angles \(\hat{\gamma}_1 = \hat{\theta}^{(1)} - \hat{\theta}^{(2)}\) and \(\hat{\gamma}_2 = \hat{\gamma}_1 + \pi\) are tested, and the one with lower MSE is selected.
12. Design Motivation: The 180° ambiguity is an inherent limitation of orientation analysis (centro-symmetric filters cannot distinguish opposite directions), but it can be resolved by a simple two-candidate comparison.

Loss & Training

The method is purely spectral and requires no training. It has no learnable parameters and is plug-and-play.

Key Experimental Results

Main Results

Synthetic images (\(N=300\), \(L=300\) superimposed Gabor atoms, von Mises distributed orientations):

Method	Angular Distance↓ (degrees)	Distribution Distance↑ (dB)
Cake wavelet	0.03 ± 0.25	94.47 ± 2.50
Ridge filter	0.06 ± 0.35	88.08 ± 2.26
Binning (baseline)	0.32 ± 0.84	50.79 ± 1.08

Real image registration:

Image	Method	Registration Error↓	Equivariance Error↓
CT scan	Cake wavelet	0.02°	0.47°
CT scan	Binning	20.00°	36.0°
Bark texture	Ridge filter	0.34°	0.36°
Bark texture	Binning	20.00°	18.00°

Ablation Study

Configuration	Key Findings	Notes
Isotropic synthetic image	Cake/Ridge distributions are nearly flat	Binning exhibits large fluctuations and rotational instability
25° oscillation synthetic image	Cake/Ridge peaks precisely align at 25°	Demonstrates the advantage of filter smoothness
Bartlett/Welch PSD	Performance degrades	Resolution loss causes degraded angular analysis
Periodogram PSD	Optimal	Lack of smoothing preserves angular resolution

Key Findings

• Cake wavelets perform best on structural images (CT registration: 0.02°), while ridge filters are slightly superior on texture images (bark: 0.34°) — the two are complementary.
• Binning registration error can reach 20° (near failure), with equivariance error up to 36°, rendering it completely unreliable in practice.
• The key advantage stems from the smooth weighted averaging of frequency-domain filters, which eliminates quantization errors at discrete bin boundaries.
• The radially symmetric window is a necessary component for rotational robustness.

Highlights & Insights

• The idea of replacing discrete bins with continuous smooth filters is remarkably elegant — it essentially resolves a discretization problem (bin boundary quantization error) using standard signal processing tools (frequency-domain filtering). The method requires no learning and is fully interpretable.
• The complementarity between cake wavelets and ridge filters provides a practical selection guide: choose the filter type based on whether the image is structural or textural in nature.

Limitations & Future Work

• Only single-resolution analysis is addressed; the approach is not extended to multi-resolution frameworks (ridgelets/curvelets/shearlets).
• The method cannot distinguish \(\theta\) from \(\theta + 180°\), requiring an additional MSE comparison step for disambiguation.
• Real image experiments include only 2 images, limiting statistical convincingness.
• No comparison with deep learning rotation-equivariant methods (e.g., E(2)-CNNs) is provided.
• Simple argmax is insufficient for complex anisotropic scenarios with multiple dominant orientations.

Related Work & Insights

• vs. Angular Binning: The primary baseline of this work. The contrast is stark: binning yields 20° registration error versus 0.02° for the proposed method.
• vs. E(2)-equivariant CNNs: Deep learning approaches require training and are black-box models, whereas the proposed purely spectral method requires no training and is fully interpretable. However, deep learning methods may offer greater flexibility in complex scenarios.

Rating

• Novelty: ⭐⭐⭐ — All components are existing tools (cake wavelets / ridge filters / PSD); the contribution lies in their systematic combination and validation of rotation equivariance.
• Experimental Thoroughness: ⭐⭐⭐ — Synthetic experiments are thorough, but only 2 real images are tested and no comparison with deep learning methods is provided.
• Writing Quality: ⭐⭐⭐⭐ — Mathematically rigorous, notation is clear, and algorithmic pseudocode is complete.
• Value: ⭐⭐⭐ — Practically useful as an orientation analysis tool, though the degree of innovation is limited.

Related Papers

• [CVPR 2026] RDFace: A Benchmark Dataset for Rare Disease Facial Image Analysis under Extreme Data Scarcity and Phenotype-Aware Synthetic Generation
• [CVPR 2026] Focus-to-Perceive Representation Learning: A Cognition-Inspired Hierarchical Framework for Endoscopic Video Analysis
• [CVPR 2026] SD-FSMIS: Adapting Stable Diffusion for Few-Shot Medical Image Segmentation
• [CVPR 2026] BiCLIP: Bidirectional and Consistent Language-Image Processing for Robust Medical Image Segmentation
• [CVPR 2026] Unlocking Multi-Site Clinical Data: A Federated Approach to Privacy-First Child Autism Behavior Analysis