Distortion-invariant object recognition using adaptive resonance theory

Classification and recognition of multiple objects under changes in position, orientation and scale are needed in practical applications such as automation of assembly lines. One of the main drawbacks in the conventional pattern recognition technique is the enormous time and computational overhead required for classification. However, the conventional techniques are well-suited for extracting the features of objects. Recently, the advantages of artificial neural networks (ANNs) of having a high degree of fault-tolerance have been used in the field of pattern classification problems. The authors have combined the advantages of both the traditional pattern recognition methodology and the neural network paradigm for the distortion-invariant object recognition. The first part of this work deals with the traditional pattern recognition techniques for the extraction of the features of objects. To extract the invariant features of objects, geometrical moment-invariant techniques are used. In the case of multiple objects, the authors do segmentation of each object before extracting the features. A neural network paradigm called the ART2-analog version of adaptive resonance theory is employed to classify objects from the extracted features