Deep learning based OCR engine for the Indus script
Computational epigraphy is an interdisciplinary area that combines computing and the study of ancient inscriptions. The main challenge or bottleneck faced in the field of epigraphical research is the lack of standardized corpora of the ancient scripts under study. Preparing such data from raw archaeological records, requires laborious human effort, expertise and a lot of time. Machine Learning has been used in the past to reduce human effort in epigraphical research, in problems such as classification and search for graphemic patterns. However, ML and in specific Deep Learning has not been applied yet, for the complementary task of corpus preparation. This talk will be focusing on how a deep learned pipeline architecture was designed to serve as an OCR (Optical Character Recognition) engine that is capable of reading the Indus script, one of the very ancient and undeciphered inscriptions of the Harappan Civilization. This pipeline takes as input, images of the undeciphered Indus script, as found in archaeological artifacts, and returns as output a string of graphemes, suitable for inclusion in a standard corpus.
This can be extended to a Full Talk too, based on the response.
- What is computational epigraphy?
- The long undeciphered Indus script
- Why still undeciphered?
- ML in the study of Indus scripts
- Need for corpus formulation
- Why deep learning?
- The design decisions in DL
- Deep features
- Transfer learning and fine tuning
- Data augmentation
- The design decisions in DL
- The deep learned pipeline architecture
- Region proposal
- Text region formulation
- Symbol segmentation
- Symbol identification
- Evaluating the pipeline’s performance
- Limitations of the pipeline
- Generalizing this architecture to other ancient inscriptions
This is Satish Palaniappan, a CS graduate from SSN CE (Sri Sivasubramaniya Nadar College of Engineering). I am currently working as a software engineer at Qube Cinema Technologies. Alongside this, I am also working as a Research Assistant under Prof. Ronojoy Adhikari at the IMSc (Institute of Mathematical Sciences), Chennai.
My work domains:
Deep Learning, Machine Learning, Computer Vision, NLP, Algorithm Design
Work at IMSc:
Currently, we are developing a deep-learned pipeline architecture that enables computers to read ancient inscriptions from images of archaeological artifacts. The working prototype of the same applied to the Indus script of the Harappan civilization and our paper titled “Deep Learning the Indus Script” arXived at: arXiv:1702.00523v1 created a buzz among the research community and in the media alike. The fully functional Indus script OCR engine will be open sourced and be available as an API based service, soon.
Work at Qube Cinemas:
- Deep learning based computer vision algorithm for mining the viewer demographics.
- Real-time adaptive selection and resource allocation algorithm for businesses.
- Research paper “Deep Learning the Indus Script”: https://arxiv.org/abs/1702.00523 (paper in arXiv, to be submitted to PLoS One)
- Related slide deck: https://rawgit.com/tpsatish95/OCR-on-Indus-Seals/master/slides/index.html#1 (Prepared during initial stages of our research)
- Related code: https://github.com/tpsatish95/OCR-on-Indus-Seals
- Press coverage:
- The Hindu: http://www.thehindu.com/sci-tech/science/chennai-team-taps-ai-to-read-indus-script/article17448690.ece
- The Verge: http://www.theverge.com/2017/1/25/14371450/indus-valley-civilization-ancient-seals-symbols-language-algorithms-ai#EQQA6r
- Times of India: http://timesofindia.indiatimes.com/city/chennai/app-may-help-decipher-indus-valley-symbols/articleshow/57281369.cms
- SBS Radio, Australia: http://www.sbs.com.au/yourlanguage/tamil/en/content/app-decipher-ancient-symbols?language=en