Fast End-to-End Deep Learning Identity Document Detection, Classification and Cropping

The growing use of Know Your Customer online services generates a massive flow of dematerialised personal Identity Documents under variable capturing conditions and qualities (e.g. webcam, smartphone, scan, or even handcrafted pdfs). IDs are designed, depending on their issuing country/model, with a specific layout (i.e. background, photo(s), fixed/variable text fields) along with various anti-fraud features (e.g. checksums, Optical Variable Devices) which are non-trivial to analyse. This paper tackles the problem of detecting, classifying, and aligning captured documents onto their reference model. This task is essential in the process of document reading and fraud verification. However, due to the high variation of capture conditions and models' layout, classical handcrafted approaches require deep knowledge of documents and hence are hard to maintain. A modular approach using a fully multi-stage deep learning based approach is proposed in this work. The proposed approach allows to accurately classify the document and estimates its quadrilateral (localization). As opposed to approaches relying on a single end-to-end network, the proposed modular framework offers more flexibility and a potential for future incremental learning. All networks used in this work are derivatives of recent state-of-the-art ones. Experiments show the superiority of the proposed approach in terms of speed while maintaining good accuracy, both on the MIDV-500 academic dataset and on an industrial based dataset compared to hand crafted solutions.