INVARIANT PATTERN-RECOGNITION FOR RANGE IMAGES USING THE PHASE FOURIER-TRANSFORM AND A NEURAL-NETWORK

A method for invariant pattern recognition of range images by means of the phase Fourier transform is introduced. The phase Fourier transform may be used for the segmentation of connected planar and quadric surfaces. The method is generalized to nonconnected planar surfaces through the use of the concept of the characteristic normal. An invariant representation under changes of position, scale, and orientation for the characteristic normals is defined. This representation is used as the input for a feedforward neural network. Examples of applications are given, and finally the method is applied to the problems of classification and occlusion.