A feature clustering algorithm for scale-space analysis of image structures

Abstract

Abstract—In describing image features it is important to consider the fact that the appearance of a feature depends on the scale or resolution at which it is observed. Existing robust image feature detectors address the issue by selecting a characteristic scale for each detected feature and subsequently describing the feature as it appears at its characteristic scale. A new method is presented for the multi-scale analysis of derivative based image features that represents a 2D image feature by its locus in scalespace. An algorithm is also presented for efficiently producing the discrete loci representations of image features through clustering features detected at multiple scales. This new method provides an entry point to potential multi-scale descriptions of image features, as well as new possibilities for feature set reduction and filtering. I.