Recognition of Objects by Supervised Neural Network Using Wavelet Features

This paper presents a simple and robust method for recognition of rotated objects by Feedforward Neural Classifier. Initially, the translation invariance is achieved after pre-processing the image. Fourier transform is then applied to each of the rotated binary edge images with 5 degrees interval. Then the 3-level Discrete Wavelet Transform (DWT) is applied to compress the Fourier coefficients. The resulting data are then divided into 16 equal non-overlapping blocks, from which the statistical descriptors are calculated and it includes mean, median, standard deviation, energy and entropy. The 80 descriptor values obtained for each rotated image are used to train the Artificial Neural Network. The simulation works are carried on the standard image dataset obtained from the Library of Amsterdam University [28]. The high classification accuracy of 97.85 % is achieved during testing.