Recognition of solid objects in images invariant to conformal transformations

Abstract

This paper extends the technique of object recognition by matching of the histograms of local orientations obtained with the structural tensor. The novel modification relies on building phase histograms in the morphological scale-space which allows inference of the intrinsic structure of the classified objects. Additionally the phase histograms are nonlinearly filtered to remove noise and improve accuracy. A matching measure has been devised to allow classification of rotated or scaled objects. It is endowed with the rotation penalty factor which allows control of preferable positions of objects. Additionally, the factor of false positives responses was lowered – a match is accepted only if it is significantly above the other two best matches. The method was tested and verified in the classification of the road signs and the static hand gestures. It showed high accuracy and very fast response.