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

Conference: CVPR2025
arXiv: 2603.11294
Code: GitHub
Area: Medical Images
Keywords: anisotropic analysis, rotation equivariance, spectral method, cake wavelet, angular registration

TL;DR

EquivAnIA, a spectral method based on cake wavelets and ridge filters, is proposed for rotation-equivariant anisotropic image analysis, demonstrating superior rotation robustness to traditional angular binning on synthetic and real images (including CT).

Background & Motivation

Anisotropic image analysis is highly prevalent in medical and scientific imaging, such as tissue texture orientation analysis in CT scans and orientation distribution detection in fibrous materials. The core task is to extract the angular profile from the 2D power spectral density (PSD) of the image to determine the dominant orientation and anisotropic characteristics of the image.

Traditional methods approximate the angular PSD \(S(\theta) = \int_0^\infty S(r,\theta) dr\) by performing angular binning on the discretized PSD over a Cartesian grid. However, due to the anisotropy of the discrete grid (e.g., more frequency points are associated with the 0° direction than the 30° direction), binning methods are highly sensitive to input rotation—rotating the same image produces different angular profiles, which severely impacts the reliability of practical applications. Particularly in angular registration tasks, rotation non-equivariance can lead to registration errors as high as 20°.

The goal of this paper is to design an anisotropic analysis method that is robust to numerical rotation, such that a rotated image produces a correspondingly rotated angular profile. The authors focus specifically on single-resolution scenarios and leave multi-resolution extensions for future work.

Method

Overall Architecture

EquivAnIA performs anisotropic analysis in three steps:

1. Window Function Preprocessing: Apply a radially symmetric, smooth window function (approximating disk support) on images with non-circular support to eliminate artifacts caused by image corners rotating in and out of the field of view, thereby enhancing the rotation robustness of the PSD estimation.

2. Directional Filter Convolution: Analyze the image using a family of directional functions \(\phi_{v,\theta}(u)\), generating filters in different orientations by rotating and translating base functions. Specifically, two types of filters are used:

   • Cake wavelet: A sector-shaped filter defined in the frequency domain, covering a specific angular range.
   • Ridge filter: A ridge-like filter defined in the frequency domain, extracting energy along a specific orientation.

3. Angular Profile Computation: The angular profile is defined as the energy response in each direction \(\rho(\theta) = \int_{\mathbb{R}^2} |c_{v,\theta}|^2 dv\), and the dominant direction is estimated by \(\eta = \arg\max_\theta \rho(\theta)\). The filters are centrally symmetric in the frequency domain, ensuring equal weighting for \(\theta\) and \(\theta+180°\).

Key Designs

• Direct use of the periodogram instead of Bartlett/Welch methods for PSD estimation, since the latter reduces noise but compromises resolution, which is unfavorable for anisotropic analysis.
• Centrally symmetric filters in the frequency domain, enabling equal-weight processing of directions \(\theta\) and \(\theta + 180°\).
• When applied to angular registration, two candidate angles \(\hat{\gamma}_1 = \hat{\theta}^{(1)} - \hat{\theta}^{(2)}\) and \(\hat{\gamma}_2 = \hat{\gamma}_1 + \pi\) are tested, and the final registration angle is chosen via minimum MSE.
• Experimental validation shows that Bartlett/Welch methods lead to worse anisotropic analysis results due to resolution loss.

Loss & Training

As a non-learning method, there is no training loss function. Evaluation is based on two metrics: angular distance (degrees) and profile distance (MSE in dB). Angular registration selects the optimal angle from the two candidates using MSE.

Key Experimental Results

Method	Angular Distance ↓	Profile Distance ↑
Cake wavelet	0.03 ± 0.25	94.47 ± 2.50
Ridge	0.06 ± 0.35	88.08 ± 2.26
Binning	0.32 ± 0.84	50.79 ± 1.08

Angular registration experiment (real images):

Image	Method	Registration Error ↓	Equivariance Error ↓
CT scan	Cake wavelet	0.02	0.47
CT scan	Ridge	0.16	0.36
CT scan	Binning	20.00	36.0
Bark texture	Ridge	0.34	0.36
Bark texture	Binning	20.00	18.00

Synthetic image experiments showcase three classes of images: isotropic images (expected constant profile), single-directional oscillating images (dominant direction of 25°), and Gabor atom superposition images (von-Mises distribution with \(\mu=60°\)).

Key Findings: Cake wavelets perform better on structural images, while ridge filters perform better on texture images; the binning method has a registration error of up to 20°, making it virtually unusable. Statistical experiments based on 300 synthetic images (each synthesized from 300 Gabor atoms, with angular parameters following a von-Mises distribution) show that cake wavelets achieve the lowest mean and lowest variance on both the angular distance and profile distance metrics.

Highlights & Insights

• Simple and elegant method: training-free, pure spectral analysis with clear theory, making it easy to implement and deploy.
• Extremely high rotation robustness: the angular error on synthetic images is only 0.03°, significantly better than the 0.32° of binning.
• Practicality is validated on real CT images (LIDC-IDRI dataset) and texture images, with registration errors as low as 0.02°.
• Two complementary filter choices are provided (cake wavelets are suitable for structural images, and ridge filters for texture images), allowing users to select according to specific scenarios.
• Binning introduces systematic biases towards grid-aligned angles (0°, 45°, 90°); the proposed method completely eliminates this issue.
• Clever design of window function preprocessing: entry and exit of corner information during rotation is eliminated by confining within a disk support.

Limitations & Future Work

• Only processes single-resolution analysis and does not extend to multi-resolution tools (ridgelet, curvelet, shearlet), which limits its analytical capacity for multi-scale structures.
• Unable to differentiate between \(\theta\) and \(\theta + 180°\), requiring additional steps (such as the Hilbert transform) for disambiguation.
• Small experimental scale (only 3 synthetic images and 2 real images), without downstream task performance validation on large-scale medical datasets.
• Lack of quantitative comparison with deep learning methods (such as rotation-equivariant CNNs or group equivariant networks).
• The robustness to noise is not fully discussed, while practical medical images usually contain considerable noise.
• Computational complexity and real-time performance are not discussed.

Related Work

• Rotation-equivariant CNNs (Lafarge et al., MedIA 2021): Learning-based method, data-driven rotation equivariance, utilized for histopathology analysis.
• Adaptive Rotated Convolution (Pu et al., ICCV 2023): Adaptively rotates convolutional kernels in object detection to process rotated objects.
• Cake wavelet (Bekkers et al., JMIV 2014): The source of the filters used in this paper, originally applied in multi-directional analysis for retinal vessel tracking.
• Ridgelet (Donoho, 2001) / Curvelet (Candes & Donoho, 2000): Classic tools for multi-resolution anisotropic analysis, while this paper focuses on single-resolution scenarios.
• Angular Difference Function (Keller et al., TPAMI 2005): Performs image registration using angular difference functions; the registration strategy in this paper is similar but based on more robust angular profile estimation.
• FFT-based Registration (Reddy & Chatterji, 1996): Classic frequency-domain registration method, for which the current paper provides a more robust front-end for angular estimation.

Rating

• Novelty: ⭐⭐⭐⭐ (Engineering improvement of classic spectral methods; limited innovation scope but with rigorous methodology)
• Experimental Thoroughness: ⭐⭐⭐⭐ (Sufficient validation on synthetic and real images, well-designed statistical experiments, but lacking large-scale application validation)
• Writing Quality: ⭐⭐⭐⭐⭐ (Clear structure, rigorous mathematics, and high-quality figures)
• Value: ⭐⭐⭐⭐ (Practical value in specific application scenarios such as CT registration and fiber analysis)

Related Papers

• [CVPR 2025] CARL: A Framework for Equivariant Image Registration
• [CVPR 2025] Boltzmann Attention Sampling for Image Analysis with Small Objects
• [CVPR 2025] Interactive Medical Image Analysis with Concept-based Similarity Reasoning
• [CVPR 2025] Multi-modal Vision Pre-training for Medical Image Analysis (BrainMVP)
• [CVPR 2025] Transformer-Based Multi-Region Segmentation and Radiomic Analysis of HR-pQCT Imaging for Osteoporosis Classification