A Hybrid Capsule Network-based Deep Learning Framework for Deciphering Ancient Scripts with Scarce Annotations: A Case Study on Phoenician Epigraphy

A hybrid capsule network-based deep learning framework for deciphering ancient scripts with scarce annotations is presented. To verify the feasibility of our proposed framework, the Phoenician epigraphy is used as a case study. A corpus of labeled data of Phoenician alphabets that covers all different styles and stages is presented. This corpus can help in contributing to the digitization process of the Phoenician culture. This dataset is preprocessed by performing conventional pre-processing techniques and then processed and augmented using a hybrid architecture of autoencoders that preserves its human-like nature. The augmented dataset is fed to a custom capsule network in order to decipher the Phoenician character and classify it into one of the 22 alphabets. Our model achieves state-of-the-art performance in recognizing handwritten characters with an overall accuracy of 0.9891 and a loss of 0.021. Therefore, our model can help develop an automated deciphering system to save epigraphists' valuable time and effort in deciphering the Phoenician epigraphy in a short period. Moreover, this work can be replicated for any other ancient scripts with minor modifications considering the systematic methodology that we proposed since it has proven its effectiveness in deciphering the Phoenician epigraphy. Our model can be employed as a transfer learning backbone for recognizing other existing alphabets which suffer from a lack of annotated data.