ANCIENT TAMIL CHARACTER RECOGNITION USING OPTIMAL THRESHOLDING WITH RESNET-CAPSNET

Ancient Tamil character recognition (TCR) is complex due to its orientation, scale and writing style varying from person to person. Further, it is a challenging task for Epigraphers. Researchers for the ancient Tamil scripts and languages are relatively less when compared to other languages. When the inscriptions are on the stone wall, it is more complex to identify the characters. This research work focuses on the recognition of different Tamil characters using automated segmentation and classification techniques. Initially, the input images are pre-processed and segmented. Here, the segmentation process is carried out by Otsu thresholding. Then, the optimal threshold value is determined by the optimization adaptive salp swarm algorithm (ASSA). Finally, the Tamil characters are recognized by the DL (deep learning) model Residual capsule network (ResNet-CapsNet). The experiment is analyzed via real-time dataset and compared with the other deep learning models. Finally, the proposed ResNet-CapsNet achieved a better overall accuracy of 91.3% and specificity of 88.3% respectively and is suitable for ancient Tamil character recognition.