On the realization and analysis of circular harmonic transforms for feature detection

Abstract

Cartesian-separable realizations of circular-harmonic decompositions for angular spectrum estimation are presented and a powerful test-statistic for rotation-invariant feature-detection in images is proposed. It is shown that pixel-domain realizations of the resulting finite impulse response (FIR) filters have a low computational complexity as a consequence of their separability and steerability. The chosen form also focuses the impulse response around the pixel-under test while ensuring adequate angular resolution after discretization. The novel test-statistic, involving angular integrals that are evaluated in the transform domain, is used to detect wedges, i.e. corners of arbitrary angle and unknown orientation, in synthetic and real imagery. Like traditional (t-and F-distributed) test statistics used in regression analysis, and unlike other rotationally invariant detectors, this similarity measure incorporates and considers uncertainty due to the limitations of finite sampling.

Real-time statistical image and video processing for remote sensing and surveillance applications

“…In [15], author focuses on the problem of feature detection in circular harmonic spectra in two-dimensional images. Examples of such a problem consists of understanding objects that do not change shape but appear differently when seen from different perspectives.…”

“…Solving such a problem is very complex and time consuming. The author in [15] proposed a deep learning-based approach for solving such problems using polarseparable responses of filter banks.…”