Phase Invariant Keypoint Detection
Anil Anthony Bharath, Nick Kingsbury · 2007
This paper introduces extensions to the complex wavelet keypoint detection paper [1], Keypoints are generated by finding local peaks in accumulated, interpolated maps of the product of magnitudes of directional complex filter responses, as in earlier work. Gradient vector fields derived from these maps are used for keypoint scale characterisation, but this is now performed so as to remove the directionality of gradient field sampling, thereby improving the stability of scale estimates. A new class of keypoints is also introduced: the circular measure (CM) keypoints, which are used to augment the locations found by the filter magnitude product (FMP) keypoints. This new keypoint class generates a higher proportion of keypoints in the interiors of objects, whilst simultaneously providing approximate object scale information, and appears appropriate for directing the attention of a vision system to the interiors of well-defined regions.