[Paper Review] A novel set of rotationally and translationally invariant features for images based on the non-commutative bispectrum

Risi Kondor|arXiv (Cornell University)|Jan 20, 2007

Blind Source Separation Techniques7 references20 citations

TL;DR

This paper introduces a novel set of rotationally and translationally invariant image features based on the non-commutative bispectrum on SO(3), achieved by projecting 2D images onto a sphere. The method computes cubic polynomial invariants in O(u^{5/2}) time, enabling effective classification with minimal data, as shown by near-perfect accuracy on rotated/translational MNIST-like digits despite only 50 training samples per class.

ABSTRACT

We propose a new set of rotationally and translationally invariant features for image or pattern recognition and classification. The new features are cubic polynomials in the pixel intensities and provide a richer representation of the original image than most existing systems of invariants. Our construction is based on the generalization of the concept of bispectrum to the three-dimensional rotation group SO(3), and a projection of the image onto the sphere.

Motivation & Objective

• Address the challenge of creating complete, non-lossy invariants for image recognition under translation and rotation.
• Overcome limitations of existing invariant methods that are either lossy or computationally expensive.
• Enable effective learning in small-data regimes by preserving discriminative information while enforcing symmetries.
• Develop a computationally tractable method for simultaneous invariance to translation and rotation in 2D images.
• Bridge abstract group-theoretic bispectrum theory to practical computer vision applications.

Proposed method

• Project a 2D image onto a sphere to map planar translations and rotations to SO(3) group actions.
• Apply the non-commutative bispectrum framework on SO(3) to generate invariants under the full Euclidean group.
• Construct features as cubic polynomials in pixel intensities derived from the bispectrum, ensuring invariance and smoothness.
• Use spherical harmonics and group representation theory to compute the bispectrum efficiently.
• Preprocess images into invariants before feeding into standard learning algorithms like SVMs.
• Set parameters via cross-validation, using L=15 and a=2 for experimental validation.

Experimental results

Research questions

• RQ1 Can a non-commutative bispectrum on SO(3) be used to generate complete, non-lossy invariants for 2D images under translation and rotation?
• RQ2 How does the computational cost of the proposed method scale with image size?
• RQ3 Can the bispectrum-based features outperform standard pixel-based representations in small-data, rotationally/translationally variant learning tasks?
• RQ4 To what extent can the method distinguish between highly similar digit classes like '6' and '9' under rotation?
• RQ5 Is the bispectrum approach viable and effective for real-world image classification with minimal data?

Key findings

• The bispectrum-based features achieved nearly 100% accuracy on easy digit pairs like '0' vs. '1' under rotation and translation.
• The method significantly outperformed baseline pixel-based SVMs, which often performed at random-guessing levels on hard cases like '8' vs. '9'.
• The algorithm successfully discriminated between '6' and '9' despite their rotational symmetry, capturing subtle handwriting differences such as pen strokes and leg shapes.
• The computational cost scales as O(u^{5/2}) time and O(u^{3/2}) memory, making it feasible for moderate-sized images.
• The features are strictly invariant and close to complete, uniquely specifying the original image up to rotation and translation.
• The method demonstrates strong generalization in low-data regimes, proving effective even with only 50 training samples per class.