Since the beginning of March, a new PhD candidate has joined ARIADNEXT research team: Timothée Neitthoffer, graduate from INSA Rennes, will conduct his PhD in our company, in collaboration with the INTUIDOC team in IRISA. He will be supervised by Bertrand Couasnon, Aurélie Lemaître and Yann Soullard in IRISA, and Ahmad Montaser Awal in ARIADNEXT.
First challenge: decrease the number of processing steps
Currently, the document analysis with IDCheck.io is done in three successive steps:
- The global structure of the document is analyzed to localize and identify the different fields (name, date of birth, photo…);
- The image is converted to a text thanks optical character recognition;
- The language is identified to make linguistic corrections or detect keywords.
The first objective of Timothée’s PhD will be to build a system able to conduct the two first steps (document analysis and character recognition) at the same time, in order to increase the efficiency of our current system.
This work will be based on the latest innovations in deep learning, especially the Attention-Based Recurrent Neural Networks developed by Xu et al. in 2015 .
The difficulty will be to take into account the specific constraints of the verification of identity documents, i.e. to be able to process and recognize a very large number of document classes, while having very few examples of each class to train the algorithms.
Second challenge: Adapting to any new class of document
The second stage of Timothée’s PhD will focus on adding a new document class to the system developed previously. The objective for the system will be to know how to adapt to this new document class. It should be able to localize the different fields and to recognize the characters with a very limited number of examples, or even ideally without any examples.
Which applications for ARIADNEXT?
This work will be carried out as part of the continuous improvement of the performance of our identity verification algorithms.
The system development and learning steps will be done according to the company’s specifications, using examples from the databases available at ARIADNEXT.
The partnership with IRISA will allow us to benefit from an expertise complementary to ours in terms of document analysis and recognition. They will also bring a different vision of the problems to be addressed, which will allow us to take a step up from our research.
A PhD at ARIADNEXT?
Your lab would like to set up a PhD in collaboration with our company? You are looking for an industrial partner for your MSCA Doctoral Network? You would like to do your PhD in ARIADNEXT? Don’t hesitate to contact us!
 Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, Yoshua Bengio, Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:2048-2057, 2015