Smart Engines: Artificial Visual Intelligence Method

The main articles are:

• Neural networks (neural networks)
• OCR - Optical Character Recognition

2023: U.S. Patent Registration for Artificial Visual Intelligence

Scientists the Russian of the company Smart Engines received for patent USA corporate development. invention allows optimising operation of neural network architectures which are used for recognitions images. The authors of the invention are a senior researcher programmer - Smart Engines, Alexander Sheshkus General Director of Smart Engines Ph.D., Vladimir Arlazarov Technical Director of Smart Engines Ph.D., Dmitry Nikolaev Professor, Corresponding Member RAS and Director of Smart to science Engines Ph.D. US Vladimir Lvovich Arlazarov 11636608 B2 is dated April 25, 2023. This was announced on May 3, 2023 by representatives of Smart Engines.

Smart Engines has registered a patent in the United States for the method of artificial visual intelligence

As reported, the inventors have proposed a neural network architecture combining blocks used in modern neural networks, with a classic real-world image analysis tool - the Hough transformation. The inventors expect that the proposed architecture will open a page in the success story of neural network technologies in computer vision. The first studies of the Hafov neural networks, already published in scientific periodicals, fully confirm this idea. The patented solution is already used in Smart Engines software products for autonomous recognition of passports, ID cards and other documents.

Neural networks extract information from examples, but it is almost impossible to teach them the immutable laws of physics or mathematics. Recent exercises of the ChatGPT network in arithmetic are indicative. When multiplying large numbers, the network correctly indicates the first and last digits of the result, and even guesses its length, but the central digits put "from the balda." Quite a funny result, because a correct solution requires billions of times less resources than those at the disposal of the neural network. The question arises: is it possible to study mathematics with examples at all? Immanuel Kant believed that a person in his knowledge relies, among other things, on a priori forms that are independent of experience. We believe that we managed to integrate into the neural network an additional a priori geometric representation underlying the laws of perspective. This allows her to build solutions to computer vision problems, such as determining the orientation of objects in space or determining her own position. told Vladimir Arlazarov, Director of Science, Smart Engines, Doctor of Technical Sciences

The most common use of the Hough transform is to find and select lines. They play an important role in the field of image processing and analysis: these are roads, and houses, and document boundaries, and lines, and X-rays that form a tomogram, and much more. But these segments are almost always not entirely straight, often noisy or only partially visible, have different lengths. Therefore, conducting a classic half-image analysis is a rather difficult task. Meanwhile, just with those problems that make it difficult to Hough-analysis of the image, neural networks cope and, moreover, there are systematic methods for solving such problems. noted Alexander Sheshkus, Senior Scientist-Programmer at Smart Engines

In 2023, the vast majority of technical vision tasks are solved using neural networks, the development of which has not raised the issue of savings for many years. At the same time, it is for vision problems that large amounts of input data are characteristic, even in trivial applications. As a result, the problem of reducing computing costs is extremely acute. Our Hough transform architecture provides competitive quality with significantly fewer learning parameters and less computing power. emphasized Dmitry Nikolaev, technical director of Smart Engines Ph.D.

This is the third patent registered by Smart Engines in the United States. In February 2023, Smart Engines scientists patented a system in the States to effectively localize and identify documents on images. In total, Smart Engines registered three patents in the United States, eight in Russia, and also created 26 useful models.