Probabilistic neural network for pattern classification
Prashanta Kumar Patra, Meghavathu S. S. Nayak, Sanjib Kumar Nayak, N.K. Gobbak · 2003
A neural network (NN) based approach for classification of images which are invariant to translation, scale and rotation is presented. The utilized network is a probabilistic neural network (PNN) classifier. The translation and scale invariant features are obtained by Fourier-modified direct Mellin transform (F-MDMT) method. The rotation invariance is obtained by Zernike moments. The classification accuracy was compared to the Bayes classifier, K-nearest neighborhood, minimum mean distance and MLP classifier. The PNN classifier gave better accuracy when compared with other methods. Similarly, F-MDMT combined with Zernike moment method was found to be better when compared with other methods like geometrical method, etc. The input data was all 49 different characters of Oriya language having each pixel size 64/spl times/64. The noise analysis was carried out and was found that the present technique was better so far as input noise is considered. This method can also be suitable for any type of 3D object recognition.