Smart Engines reported on the scientific activity in the AI field in 2023
In 2023, Smart Engines reported significant scientific achievements. The organization’s researchers published 41 science papers in international academic editions and obtained 6 patents. Additionally, they presented over 20 reports at various conferences and made several scientific discoveries.
Four scientific articles were published in the first quartile (Q1) journals of the Web of Science (WoS), and 11 articles were published in journals classified as Q1 according to the Scopus database. Two employees defended their dissertations for the D.Sc, and the other two for the degree of Ph.D.
Throughout the year of 2023, the company’s scientists participated in three international conferences. One of them is the 17th International Conference on Document Analysis and Recognition (ICDAR 2023), which is considered to be the leading event (CORE A) in the field of document recognition worldwide. The researchers of Smart Engines also addressed the London International Meeting (LIM 2023), and the 16th International Conference on Machine Vision (ICMV 2023). Smart Engines scientists presented more than 20 reports at these conferences in total.
In August 2023 researchers of Smart Engines presented a public annotated corpus (dataset) of the document images MIDV-Holo, by means of which inventors from all over the world are able to teach their AI algorithms to combat presentation attacks and rebroadcast attacks committed.
One of the outstanding scientific results was highlighted in the article published in the Expert Systems with Applications journal (Q1 WoS). There was presented a unique protocol for automated neural network-based detection of coronavirus in chest CT scans, which reduces the dose of X-ray radiation without significant impact on the quality of diagnostics.
In April 2023, the Mathematics journal (Q1 WoS) published our employees’ new study, which introduced an innovative, per-neuron approach to training quantized neural networks with low bit-widths. Such neural networks operate much more rapidly than their traditional counterparts, and the proposed approach is simple to implement, achieving the same quality as using more complex methods.
Moreover, our specialists invented a new versatile architecture for robust hypothesis estimation PESAC (Parallel Efficient Sample Consensus), allowing to implement a wide range of modern algorithms and achieve a 2–3 times acceleration of document localization and tracking (relative to other RANSAC architectures) on CPUs due to efficient data organization and parallel processing. This result was published in the interdisciplinary journal IEEE Access (Q1 Scopus).
This year the company’s scientists have also presented a work devoted to the rapid detection of machine-readable zones (MRZ) for document recognition on mobile devices. The detection algorithm is based on a proprietary neural network architecture that identifies MRZ in just 16 milliseconds on an iPhone SE 2020. In other words, on a mobile processor released in 2019, the detection rate reaches 62 frames per second. The described technology is already used for recognizing MRZ in the Smart ID Engine system, which allows to enter passport data 20 times faster and twice as accurately as a qualified operator.
Six inventions received patents in the USA. The company protected several of its key technologies with patents, including methods for document identification, frame integration in video streams, stopping text recognition in video streams, and using the Hough transform in neural networks. Another patent was granted for a key invention in the field of computer tomography. The patented solutions are already used in Smart Engines’ software products for autonomous recognition of passports, ID cards, and other documents.