Recognition of objects given by collections of multichannel images
M. M. Lange, D. Yu. Stepanov · Pattern Recognition and Image Analysis · 2014
This paper proposes a multiclass metric classifier of composite objects given by collections of multichannel images that are generated by various sources. The classifier is constructed in a space of multilayer tree-structured representations of the objects and based on making a decision by weighted voting of template objects. In the set of the object representations, we define a family of embedded measures, which provides a scheme of fusing the sources and channels by the general weighted measure. A computational gain of a proposed guided search algorithm as compared with an exhaustive search decision algorithm is estimated analytically. An efficiency of the classifier is demonstrated by experimental estimates of recognition error rates for biometric composite objects that are produced by a couple of sources which generate grayscale images of signatures and color images of faces. The recognition error rates over the individual channels of the sources, as well as over the couple of the sources, are given. An advantage of the fusion scheme by the general weighted measure in relation to the known fusion scheme by voting decisions over individual channels is shown.