Handwritten Character Recognition from Ancient Palm Leaves using Gabor based MultiLayer Architecture: GMA

摘要

Feature extraction plays the key role in pattern recognition systems. With the invent of Deep learning algorithms it is believed that the importance of feature extraction methods has been reduced. But the cost of implementing deep algorithms is very high. Deep neural networks rely on GPU architecture and requires large amount of data for achieving high recognition efficiency. It is computationally expensive to train the deep architecture and more over the learning procedure and factors for training is not easy to realize. Therefore having inspired by the structure of Deep Convolutional Neural Network a new feature extraction method based on conventional feature extraction system for recognition is proposed. In this method a multilayer architecture is designed with convolution layer based on gabor filter and classification layer based on Artificial Neural Network. Only the classification layer is subjected to learning by backpropogation and all other layers acts as the part of feature extraction system. The input image without applying any pre-processing can be subjected to the proposed system which in turn predicts the class of image as output. The proposed method is compared with some of the existing efficient feature extraction methods like discrete meyer wavelet, zernike moment, curvelet, legendre moments, gaussian hermite(GH) moment and Histogram of gradient(HOG). The recognition efficiency produced by the method without applying any pre-processing on input images is much higher than existing efficient feature extraction methods with preprocessing applied. The proposed method works effectively invariant to noise, translation and rotation. Experimental analyses were carried out in two datasets. First dataset is the standard HPL dataset of isolated Tamil characters. The second dataset consists of Grantha characters extracted from ancient palm leaves.