The core aspiration of this proposed work is to classify Tamil characters inscribed in the palm leaf manuscript using an Artificial Neural Network. Tamil palm leaf manuscript characters in the form of images were processed and segmented using contour-based convex hull bounding box segmentation. The segmented characters were transformed into two forms: Binary Coded Value and the Gray-Level Co-occurrence Matrix (GLCM) feature. The features extracted from the segmented characters were trained by the proposed method of the Modified Adaptive Backpropagation Network (MABPN) algorithm with Shannon activation function. Weight initialization plays an important role in the Backpropagation Neural Network, and hence Nguyen-Widrow weight initialization was introduced to initialize the weights instead of random weight initialization in the proposed method. The models evaluated are MABPN with Shannon activation function using Nguyen-Widrow weight initialization in two forms of input: Binary Coded Value and GLCM feature extracted values. The proposed method with GLCM features as input gave a promising result over binary coded transform.

Offline handwritten identification of characters is a core problem in pattern matching. The main challenge for researchers in the identification of handwritten characters is inscribing individual styles. Tamil mantras identification is a challenging job due to many missing features in the mantras ' complex structure. The problem at hand is to break traditional hand-designed features. A new venture has been undertaken to automatically extract the complex features for recognition and classification from the complex structure bypassing the individual Tamil characters into the Convolutional Neural Network, a special type of deep learning network. The best convolutional model was chosen to improve efficiency by comparing different convolutional models that vary in activation functions, classifiers, and pooling functions. Principal Component Analysis (PCA) was used to select the top n eigenvectors from the image for better efficiency. So with the above trained best model with PCA for independent Tamil character images, handwritten Tamil fonts in the slogans (a group of characters) have been well recognized.

Character classification in the handwritten Tamil palm-leaf manuscript is more challenging than the other document character classification due to degradation and ancient characters in the palm-leaf manuscript. In this work, RBF (Radial Basis Function) network and CART (Classification and Regression Tree) were used to classify the Tamil palm leaf segmented characters. This work consists of two phases: In the first phase, the scanned Tamil palm leaf images were preprocessed by converting them into a grayscale image and then the images were allowed to remove noise using a median filter. In the second phase, GLCM (Gray Level Co-occurrence Matrix) feature extraction method was used to extract the statistical features from the segmented characters and these features were used to train the RBF network and CART algorithm. For the RBF network, Nguyen-Widrow weight initialization technique was used to generate the weight instead of random initialization. The dataset used in this work is Kuzhanthai Pini Maruthuvam (Medicine for child-related disease). By comparing RBF using Nguyen-Widrow method with CART algorithm, RBF yields promising result of 98.4% of accuracy whereas CART produced 98.8% of accuracy for character classification. The digitization of the Tamil palm-leaf manuscript will preserve the historical secrets, traditional medicine to cure disease, healthy lifestyle, etc. It can be used in the archeological department and Tamil libraries having a palm leaf script to preserve the manuscript from degrading.