An End-to-End OCR-Free Solution For Identity Document Information Extraction

Efficiently verifying the customer’s identity is essential for ensuring the security of online transactions. Such verification is mainly accomplished according to the widespread Know Your Client (KYC) protocol, which relies on a self-identification handled by the user, typically providing personal data from their identification document (ID). In the current digital communication generation, such a task is usually performed by uploading a digital copy of the document by scanning or taking a real-time picture from a personal device such as a smartphone or tablet. Such activity, which usually involves manual data entry, is time-consuming and prone to human errors. Document Understanding (DU) has emerged as a crucial factor for automating data extraction in this scenario. One of its main challenges is the scarcity of data, which is barely available for security and privacy concerns. To this end, this paper proposes a solution that takes advantage of recent advances in DU to devise an innovative strategy for Identity Document Recognition (IDR), i.e., the task aimed at automatically understanding, extracting, and transcribing the fields of an ID card. We devised a two-stage approach based on fine-tuning a pre-trained model on synthetic and real-world data. We also developed a dedicated synthetic data generation tool to support the IDR process. Experimental results demonstrated the effectiveness of our methodology.