Image features based on local Hough transforms

Abstract

A new method of building local image features is proposed. The features are represented by various shapes (patterns) that can be approximated using Hough transforms. However, the transforms are applied locally (to the current content of a scanning window) so that the shape's location is fixed at the current window's position. Thus, the parameter-space dimensionality can be reduced by two (compared to globally computed Hough transforms) and the transforms can be effectively applied to more complex shapes. More importantly, shapes can be decomposed (two decomposition schemes are proposed) so that the overall complexity of the shapes used as features can be very high. The proposed feature-building scheme is scale-invariant (if scale is a dimension of the parameter space) subject only to diameters of scanning windows. © 2009 Springer Berlin Heidelberg.