On using sign method for 3D images recognition and classification

Nowadays different ways of 3D scene creating is used in different branches of information processing. These are 3D modeling in systems of design, and digitalizing of real 3D objects using 3D scanners or from set of photos using stereo matching methods. Huge number of objects and scenes, constructed by human often demand sorting and classification. In robotics there is a problem of objects recognition in process of robot motion. Analysis of 3D scanner data, which may contain information about large indoors, raise a problem of automatic scene objects recognition. In this work we propose method of computing measure of proximity between two shapes of 3D objects. Results of experiments with handmade objects are given also. This method is invariant to moving and slight deformation of objects. Our approach differs from classic methods, based on pattern matching and Fourier specters analysis. Namely, objects are represented as a structure calling semantic net, which contains information about statistical connections between small pieces of object. Thus, in case of object light deformation our method gives good results. We have got good results applying the same method to speech and 2D images recognition. It should be noticed that in all applications the same algorithms of semantic net comparison were used. Differences are only in ways of transforming object into semantic net.