부산대학교 도서관

Recognizing appropriate ancient Tamil characters in epigraphy will provide detailed insights into past civilizations. The main challenges in text restoration are perspective distortion, variation in the shape of texts and non-availability of reference images. Most of the inscriptions contain a lot of noises which arise due to corrosion and other environmental factors. Due to the evolution of the language, the clustering of characters results in false positives. The conventional neural networks fail to address the above-mentioned challenges simultaneously. This article aims to assist the epigraphers through image preprocessing for denoising, feature extraction, and classification using Ancient Tamil Character Recognition in Inscription - 21 (ATCRI-21) based convolutional neural network. The preprocessing stage includes grayscale conversion, noise removal and binarization with Otsu's thresholding. Image denoising is achieved by clustering the noise due to corrosion, spots and scratches into a single cluster and the noise in the image is rejected. The character recognition is carried out through ATCRI-21 that includes three sets of a five-layer block (two convolution layers, a Rectified Linear Unit (ReLU) layer, two max-pooling layers and a dropout layer) followed by two sets of a two-layer block (a dense layer and a ReLU layer) followed by the output layer and softmax activation layer. The inclusion of regularisation layer, prevents overfitting of data. This helps us to train the network in less time. The outcome of this model is trained, validated and tested through real-time images collected from Brihadeeshwar temple, Thanjavur. The proposed model displays an validation accuracy of 80%. The outcome of this work will correlate the history of the past civilization. The proposed model unlocks a cooperative potential between epigraphers and artificial intelligence and impacts our study of important periods of history.