A two-stage neural net for segmentation of range images

A two-stage neural network is proposed for segmentation of range images. Emphasis is placed on a neural network (NN) based system that integrates edge and surface information to generate robust surface maps in the range data. The proposed architecture has two stages. The first stage extracts the surface information through self-learning least-squares surface fitting along a set of nonorthogonal basis functions. Daugman´s projection NN stage locally computes the surface normals in the image. In the second stage, the surface and edge information complete with each other to perform region growing. The edge information is obtained using a set of Zernike moment-based operators. Kohonen´s self-organizing NN is used to implement the competitive region-growing. Experimental results with real images demonstrate the effectiveness of the proposed NN architecture