OCR Studio, a developer of ID scanning solutions, unveiled IIRDoc-Net – an
advanced ultra-lightweight neural network architecture that is able to recognize IDs
with 32% fewer operations than existing state-of-the-art models. The new AI
technology is available on mobile phones, tablets, and desktops, and is already used
by the company’s clients from banking, retail, and healthcare. The invention was
presented at the 18th International Conference on Machine Vision (ICMV 2025) in
Paris.
Automated identity document recognition is widely used in areas such as banking,
logistics, and border control. Today’s shift toward mobile platforms demands these
technologies to be performed even on low-budget devices with limited computational
resources. At the same time recognition models must have an extended receptive
field to keep the calculations high-accurate and to deal with document images’
projective distortions, noise, motion blur, and poor lighting. However, most of the
models with wide receptive fields are too heavy to deliver a real-time document
analysis on a mobile device which is crucial for modern online services.
OCR Studio’s solution addresses this problem by integrating a block of learnable
Infinite Impulse Response (IIR) filters. These filters allow the receptive field to be
adjusted by incorporating both current inputs and prior outputs, which is typically not
implied even by the state-of-the-art methods. Evaluation on MIDV-500 and MIDV-
2019 datasets demonstrated a 32% reduction in computation needed for ID
recognition compared to best competing approaches. Actual technology has been
implemented in the latest version of OCR ID-scan that is effectively used by
government, fintech, HR, and other services for document-processing automation.
A paper about the IIRDoc-Net technology, unveiled by the OCR Studio team, will be
published in the electronic proceedings of the International Society for Optics and
Photonics (SPIE), marking the team’s debut at a major international scientific
conference of such level. Articles presented at the ICMV are indexed in Web of
Science Conference Proceedings Citation Index-Science, Scopus, Ei Compendex,
and other research databases.
