Polynomial-based rotation invariant features
This addresses the difficulty of rotation invariance in image recognition and shape similarity evaluation, offering a more efficient alternative to costly pairwise alignment methods.
The paper tackles the problem of handling unknown rotations in machine learning by proposing a method to construct arbitrarily large sets of rotation invariants from polynomials, achieving up to O(n^D) independent invariants compared to O(D) from standard approaches.
One of basic difficulties of machine learning is handling unknown rotations of objects, for example in image recognition. A related problem is evaluation of similarity of shapes, for example of two chemical molecules, for which direct approach requires costly pairwise rotation alignment and comparison. Rotation invariants are useful tools for such purposes, allowing to extract features describing shape up to rotation, which can be used for example to search for similar rotated patterns, or fast evaluation of similarity of shapes e.g. for virtual screening, or machine learning including features directly describing shape. A standard approach are rotationally invariant cylindrical or spherical harmonics, which can be seen as based on polynomials on sphere, however, they provide very few invariants - only one per degree of polynomial. There will be discussed a general approach to construct arbitrarily large sets of rotation invariants of polynomials, for degree $D$ in $\mathbb{R}^n$ up to $O(n^D)$ independent invariants instead of $O(D)$ offered by standard approaches, possibly also a complete set: providing not only necessary, but also sufficient condition for differing only by rotation (and reflectional symmetry).