INTELLIGENT SYSTEMS AND TECHNOLOGIES

Computationally efficient detection of the region of interest of the Russian Passport in the image

A. V. Gayer<sup>ab

a</sup> Federal Research Center "Computer Science and Control" of Russian Academy of Sciences, Moscow, Russia
^b Smart Engines Service LLC, Moscow, Russia

Abstract: The article addresses the task of localizing a Russian Federation citizen's passport in photographs where the document occupies a small portion of the frame. This problem is particularly relevant for remote verification systems that require users to upload a selfie with their passport. The small scale complicates document recognition and localization due to lower resolution. To improve localization accuracy, an ultra-lightweight neural network model, YOLO-Passport, is proposed for passport region localization, reducing the problem to a fixed document scale. Compared to compact YOLO detectors, YOLO-Passport has an order of magnitude fewer operations and parameters. The proposed approach increased the detection recall of Russian passports from 91.6% to 97.4%. The model's inference time on a CPU is 3 ms, and its size in 8-bit format is only 340 KB, making it efficient for deployment in industrial systems and WASM-based web applications.

Keywords: document recognition, deep learning, object detection, YOLO.

DOI: 10.14357/20718632250301

Bibliographic databases:

•