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

Conference: CVPR 2026
arXiv: 2603.11294
Code: github.com/jscanvic/Anisotropic-Analysis
Area: Signal Processing / Image Analysis
Keywords: rotation equivariance, anisotropic analysis, spectral method, cake wavelet, angular registration

TL;DR

This paper proposes EquivAnIA, a spectral method that computes angular energy distributions via Cake wavelets and Ridge filters in the Fourier domain, achieving strictly numerically rotation-equivariant anisotropic image analysis. The method substantially outperforms conventional angular PSD binning approaches on both synthetic and real images.

Background & Motivation

Background: Extracting directional information from images is widely used in medical imaging (e.g., CT vessel orientation identification, fibrous tissue analysis) and scientific imaging. The core tool is the angular power spectral density (angular PSD), which integrates the 2D PSD over angles to obtain the energy distribution across orientations.

Limitations of Prior Work: In practice, the PSD is estimated on a Cartesian grid, and the angular PSD is approximated via angular binning. Binning is not numerically equivariant under rotation — rotating an image does not cause the analysis result to shift accordingly. Grid-aligned directions (0°, 45°, 90°, etc.) contain more frequency points in their bins, introducing systematic bias. This is particularly problematic in applications requiring precise angular estimation, such as angular image registration.

Key Challenge: The discretization of the Cartesian frequency grid breaks the continuous rotational symmetry, causing binning-based methods to produce angle-dependent artifacts.

Goal: Design a rotation-equivariant analysis method satisfying \(f(R_\alpha I) = \text{shift}_\alpha(f(I))\) — applying a rotation \(\alpha\) to the image should cause the angular distribution to shift exactly by \(\alpha\).

Method

Overall Architecture

Input image → radially symmetric window function for boundary handling → DFT to frequency domain → directional filters (Cake wavelet or Ridge filter) computing weighted energy at each angle \(\theta\) → output angular energy distribution \(\rho(\theta)\) → used for dominant direction estimation \(\eta = \arg\max_\theta \rho(\theta)\) or angular registration.

Key Designs

1. Cake Wavelet Directional Filtering:

   • Function: Partitions the frequency domain into \(K\) overlapping "cake-slice" sector filters, each covering an angular range of \(2\pi/K\), and computes the weighted energy in each direction.
   • Mechanism: Defines a family of directional functions \(\phi_{v,\theta}(u) = \phi(R_\theta^{-1}(u-v))\), with angular energy \(\rho(\theta) = \int_{\mathbb{R}^2} |c_{v,\theta}|^2 dv\), where \(c_{v,\theta}\) are analysis coefficients. Cake wavelets are parameterized directly in the frequency domain, guaranteeing rotation equivariance.
   • Design Motivation: Unlike binning, Cake wavelets apply smooth weighted averaging over each angle rather than discrete bin counting, eliminating the bias at grid-aligned orientations. The filter shapes co-rotate under rotation, providing theoretical equivariance guarantees.

2. Ridge Filter:

   • Function: Uses anisotropic Gaussian filters to enhance "ridge-like" structural responses along specific directions.
   • Mechanism: Parameterized in the frequency domain as elongated Gaussian windows along a given direction, providing superior directional selectivity for elongated structures such as vessels and fibers.
   • Design Motivation: Cake wavelets offer better robustness for general textures, while Ridge filters are more sensitive to elongated structures. The two are complementary, and users may select based on image content type.

3. Radially Symmetric Window Preprocessing:

   • Function: Applies a smooth approximately circular support window to non-circular images, discarding corner information.
   • Mechanism: When an image is rotated, information in the corners enters and exits the rectangular boundary, introducing non-equivariant errors. A radially symmetric window ensures that the region participating in analysis remains invariant under rotation.
   • Design Motivation: This is critical for achieving equivariance in discrete implementations — rotation preserves the analysis region only when a circular support is used.

Angular Image Registration Algorithm

Given two rotated copies \(x^{(1)}, x^{(2)}\): (1) compute angular energy distributions \(\rho^{(1)}(\theta), \rho^{(2)}(\theta)\) respectively; (2) estimate dominant directions \(\hat{\theta}^{(k)} = \arg\max \rho^{(k)}(\theta)\); (3) accounting for 180° ambiguity, test two candidate angles \(\hat{\gamma}_1 = \hat{\theta}^{(1)} - \hat{\theta}^{(2)}\) and \(\hat{\gamma}_2 = \hat{\gamma}_1 + \pi\); (4) select the candidate minimizing MSE as the registration result.

Key Experimental Results

Main Results (Synthetic Images — 300 Random Gabor Atom Images)

Method	Angular Distance↓ (°)	Profile Distance↑ (dB)
Cake Wavelet (Ours)	0.03 ± 0.25	94.47 ± 2.50
Ridge (Ours)	0.06 ± 0.35	88.08 ± 2.26
Binning (Baseline)	0.32 ± 0.84	50.79 ± 1.08

Ablation Study (Real Image Registration)

Image	Method	Registration Error↓ (°)	Equivariance Error↓ (°)
CT Scan	Cake Wavelet	0.02	0.47
CT Scan	Ridge	0.16	0.38
CT Scan	Binning	3.13	2.99
Bark Texture	Cake Wavelet	0.45	1.00
Bark Texture	Ridge	0.04	0.04
Bark Texture	Binning	7.88	6.76

Key Findings

• Cake wavelets maintain consistently low error across all rotation angles, whereas binning degrades significantly at non-grid-aligned angles.
• Ridge filters outperform on texture images (bark registration error: 0.04° vs. 0.45° for Cake), while Cake wavelets outperform on structural images (CT: 0.02° vs. 0.16° for Ridge).
• The equivariance error of binning is 10–100× larger than that of EquivAnIA.
• After 90° rotation, EquivAnIA's angular distribution shifts precisely, while binning exhibits visible bias.

Highlights & Insights

• Mathematical Elegance: The paper extends rotation equivariance from a continuous-domain theoretical guarantee to a discrete numerical implementation, addressing a theoretical gap in conventional methods.
• As a purely signal-processing approach requiring no training, the method is computationally efficient and applicable to any scenario requiring directional analysis.
• The complementarity between Cake wavelets and Ridge filters offers practical guidance: Cake for structural content, Ridge for textural content.
• The angular registration application is simple and effective, and can serve as initialization for more complex registration pipelines.

Limitations & Future Work

• Only single-resolution analysis is addressed; equivariance for multi-resolution extensions (ridgelets, curvelets, shearlets) is left for future work.
• Angular estimation cannot distinguish \(\theta\) from \(\theta + 180°\); additional processing such as the Hilbert transform is required to resolve this ambiguity.
• Real-image experiments involve only 2 images, lacking large-scale quantitative evaluation.
• No comparison is made against deep learning-based orientation estimation methods (e.g., SteerableCNN).

Related Work & Insights

• Classical Spectral Analysis: The angular PSD is a conventional tool; EquivAnIA is its equivariant reformulation.
• Steerable Filters / Equivariant CNNs: Rotational equivariance has been extensively studied in deep learning (e.g., E(2)-CNNs); EquivAnIA provides a complementary signal-processing perspective.
• Insight: Carefully designed classical signal processing methods can still surpass general-purpose deep learning approaches in specific scenarios, particularly with respect to mathematical properties requiring strict theoretical guarantees.

Rating

• Novelty: ⭐⭐⭐ The core idea (replacing binning with directional filters) is not entirely new, but the systematic experimental validation and registration application are valuable contributions.
• Experimental Thoroughness: ⭐⭐⭐ Synthetic image experiments are detailed, but real-image evaluation covers only 2 images without large-scale assessment.
• Writing Quality: ⭐⭐⭐⭐ Mathematical derivations are rigorous and clear, with a complete logical chain from problem statement to methodology.
• Value: ⭐⭐⭐ The method is valuable for specific applications requiring rotation-invariant directional analysis, though its scope of applicability is relatively narrow.

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