A Neural Architecture for 2-D and 3-D Vision

Abstract

This paper presents a model-based neural vision system. Scenes are described in terms of shape primitives (line segments derived from edges in the scenes) and their relational structure. The neural network matches the primitives in the scene to the primitives in a model base by finding the best agreement between primitives and their relational structure under the constraint that at most one primitive in the model base should be assigned to a primitive in the scene. The quality of the solutions and the convergence speed were both improved by using mean field approximations. This approach was tested in 2-D and in 3-D object recognition. In the 2-D problem, the recognition is independent of position, orientation, size and small perspective distortions of the objects. In the 3-D problem, stereo images are used to generate a 3-D description of the scene which is then matched against objects in a model base.