Knowledge Integration Inside Multitask Network for Analysis of Unseen ID Types.

Identity Document recognition is a key step in Know Your Customer applications where identity documents (IDs) are verified. IDs belonging to the same type share the same field structure called template. Traditional ID pipelines leverage this template to guide the localisation of the fields and then the text recognition. However, they have to be tuned to the different templates to correctly perform on those. Thus, such pipelines can not be directly used on new types of IDs. In this work, we address the task of text localisation and recognition in the context of new document types, where only the template is available with no labeled samples from the new ID type. To that end, we propose the use of Context Blocks (CB) performing template self-attention to guide the features of the network by the template. We propose three ways to leverage CB in a multitask architecture. To evaluate our approach, we design a new public task for the MIDV2020 database from rectified in-the-wild photos. Our method achieves the best results for two datasets including an industrial one composed of real examples.