Hierarchical rule-based neural network for multi-object classification using invariant features

Feature extraction techniques play a vital part in pattern recognition applications. In order to achieve the best performance in a particular classification problem, the most appropriate feature extractor for the problem is pursued. In this paper, a Pseudo-Zernike Moments based model is used as the feature extractor due to its reliability in illumination and rotation invariant multi-class object classification. A Hierarchical Rule-Based Neural Networks (HRB-NN) approach is proposed to classify multi-class data using hierarchical classification based on similarity measures between different classes. HRB-NN performance is compared to Nearest Neighbor and Bayesian classifiers. For implementation, a database of 960 images (640 training, 320 testing) for 8 different objects is used. The proposed method was able to classify the given data without any failure by giving the best performance outperforming the other chosen classifiers.