Wavelet Analysis based Handwritten Script Classification using Auto Encoders
N. Shobha Rani, M. Keerthi · 2020
Handwritten script recognition and classification is one of the unsolved challenges in an area of offline character recognition. In this paper, handwritten script classification is performed on different South Indian scripts Kannada, Telugu, Tamil, and Malayalam along with Devanagari and Roman Scripts. The proposed model for handwritten script classification initiates with a knowledge repository of word-level and character-level handwritten data samples that are collected from more than 500 writers and extracted using, projection histogram features of documents. Datasets are extracted at the character and word level separately and forwarded for wavelet feature analysis, using DB-3, DB-4, Haar, and Symlet, and the feature vectors are modeled for character level and word level datasets. These features are used for classification which is conducted separately for character and word level wavelet features. The features are employed separately for wavelets DB3, DB4, Haar, and Symlet during classification. A comparative analysis is conducted using a directional gradient feature which has shown dominant results towards printed script classification in literature. Classification is conducted using Autoencoders for both wavelet features and directional gradient features, along with that, the efficiency of conventional classifiers like support vector machines, KNN, and boosted trees is also tested with wavelet features. In the proposed method, Symlet features provide efficient performance with Autoencoders compared to other features with 70.61% accuracy.