SPIE, Vol. 3516 , 1998_____ 23-th International Congress on High Speed Photography and Photonics_____ Optimization of Informative Components for 3-D Object Recognition Alexander B. Murynin, Victor D. Kuznetsov, Ivan A. Matveev Computer Centre of Russian Academy of Sciences, 40 Vavilova, Moscow, Russia, 117967 E-mail: murynin@ccas.ru Phone: 7(095)9307237 ABSTRACT Work presented suggests a combined informational space and decision rule for recognition of 3-D objects. The informational space consists of heterogeneous sets of features (i.e. belonging to different spaces), that are object images, images of certain object features and 3-D object surface representation. Decision rule for recognition in this combined space is proposed. The method was tested on a database of human face stereo-images and gave a significant improvement of reliability of automatic recognition system. Keywords: computer vision, image recognition, principal component analysis. 1. INTRODUCTION The problem of automatic real time recognition of 3-D objects is urgent for high performance computer vision systems. One of the fields of recognition is face recognition. 1 Principal components analysis is a widely used method. Usually, this method is implemented to informational space constructed on the basis of image pixel values (signal or brightness). Meanwhile, fast and reliable stereo-reconstruction techniques have been developed. 2 Hence the problem is to implement common-use algorithms of recognition to the informational space that combines 2-D and 3-D data. Classic problem of image recognition may be defined as follows. Objects of recognition are given as images and image is a set of features or vector in a multidimensional space. Training set of images where for every image it’s object is known is inputted in the algorithm. In other words, if all objects are divided into classes, one will know which class every image belongs to. On the basis of this information, algorithm of recognition is generated. Algorithm should define which class (or classes) new input image belongs to. Images usually consist of raw and bulk sets of features. It can be raster photo-image, digitized audio signal, etc. Usually this data is inconvenient for direct use in recognition algorithm. Informative features optimization problem occurs on a stage of image preprocessing, therefore before recognition process itself. Preprocessing is intended to solve two main problems: clusterization and features selection. Clusterization is a grouping of vectors representing objects of the same class. The results of clusterization are minimizing inner-class distances and maximizing inter-class distances. Effective features selection may dramatically reduce number of dimensions of image space. Obviously this results in informativity loss, however optimal selection of features minimizes these losses and even may improve recognition results in the case of limited time and computational cost. The point is while reduction of number of features and accordingly reduction of processing time are substantial, informativity loss is very low. Optimal features selection should maximize reduction of number of features and minimize informativity loss. Possible optimization methods can be based on signal filtering, using a priori knowledge of objects of recognition, statistical analysis, etc. Principal Components Analysis (PCA) is an effective statistical method of informative features optimization. It is widely applied in image recognition, particularly in face recognition. 3,4,5 This method has a characteristics of clusterization and optimal feature selection, because it minimizes original image reconstruction error. In this work an extension of PCA is proposed. This extension can be useful for compound images consisting of images