Object recognition using invariant object boundary representations and neural network models

Abstract

Object recognition is an essential part of any high-level computer vision system. In this paper, several approaches for classifying two-dimensional objects which are based on the use of both invariant boundary transformations and artificial neural networks (ANNs) were implemented and compared. Specifically, the centroidal profile, the cumulative angular and the curvature representations were used. Two different ANN learning approaches were considered. The first involved supervised learning while the second involved unsupervised. In particular, the multilayer ANN trained with the predict back-propagation rule and the Kohonen ANN were utilized. Implementation issues, simulation results and comparisons show the strengths and weakness of each approach, especially when noisy and distorted objects were used for recognition.