Recognition of Characters using PCE based Convolutional LSTM Networks from Palaeographic Writings

Abstract

The historic paleographic writings that contributes to cultural heritage of India were inscribed on various materials such as stone inscriptions, rock carving, palm manuscripts, pots, coins, copper plates etc. Archaeological departments throughout the world have undertaken massive digitization projects to digitize the historical contents. But it is highly complicated as it involves images with complex backgrounds, noises and various illumination conditions. The paleographic writings are camera captured and processed for recognition of characters. A character recognition system is an inevitable tool to offer global visibility to the paleographic writings. Automatic character recognition is a challenging problem as in the proposed work it needs a cautious blend of image enhancement, patch extraction, feature extraction, classification and recognition techniques. This involves extracting the sequence of image patches and feature vector of the patches using Convolutional Neural Network and feeding the feature vectors using attention mechanism to recognize the character with LSTM model. As paleographic writings have lengthy sequence of characters which requires special attention during character recognition. The proposed work is an attempt to identify and recognize the historical Tamil paleographic writings by extracting the sequence of patches from the image and feeding them into a CNN-LSTM framework. The proposed method mainly consists of pre-processing, feature extraction, and character-level recognition. The LSTM network is built and the sequence of feature vectors is fed to the network and trained. The sequence of characters is recognized. The performance of the proposed method recorded an character recognition accuracy of 97.9%.