Face and Object Recognition and Detection Using Colour Vector Quantisation

In this paper we present an approach to face and object detection and recognition based on an extension of the contentbased image retrieval method of Lu and Teng (1999). The method applies vector quantisation (VQ) compression to the image stream and uses Mahalonobis weighted Euclidean distance between VQ histograms as the measure of image similarity. This distance measure retains both colour and spatial feature information but has the useful property of being relatively insensitive to changes in scale and rotation. The method is applied to real images for face recognition and face detection applications. Tracking and object detection can be coded relatively efficiently due to the data reduction afforded by VQ compression of the data stream. Additional computational efficiency is obtained through a variation of the tree structured fast VQ algorithm also presented here.