A robust-invariant pattern recognition model using Fuzzy ART

The purpose of this paper is to present a pattern recognition model that possesses both robust and invariant properties. A 'robust and invariant' concept is defined as follows: First, the pattern recognition model can recognize the objects that are translated, scaled, and rotated. Second, the system must have strong resistance to noise. Finally, the completely learned system can recognize new objects in other categories without changing any parameters of model. A new invariant vector named fuzzy-invariant vector (FIV) is introduced to the input data model. For computing FIV, known technologies, such as contouring, spectral analysis, fuzzy number, and confidence interval are used. Fuzzy ART is used as a classification model and the vigilance level of Fuzzy ART influences its performance. To improve its performance, a method that finds the appropriate vigilance range is used. To verify the performance of this model, three kinds of experiments were conducted such as learning and testing for given patterns, testing adaptability for new patterns, and comparing FIV with invariant vector (IV). Images of 11 flights and 10 tools were used in these experiments. Experimental results revealed two facts: First, this model has a recognition rate higher than 99% when an object with noise is translated, scaled, and rotated. Second, the completely learned model can recognize new patterns that have not yet been learned, and can do so at a recognition rate of over 94%. FIV gives an invariant to the model and reduces the effects of noise. Fuzzy ART is non-supervised neural network for solving the stability-plasticity dilemma, and the combined effects of FIV and Fuzzy ART yields robustness.