19.01.2024 г.

Smart Engines trained AI to see hidden spaces and patented its technology in the USA

Smart Engines has received an American patent for another one of its proprietary inventions. With its help, AI can restore the parameters of three-dimensional space from a two-dimensional image. Smart Engines scientists have proposed using Hough Transforms as layers of a neural network, making it better at computer vision tasks, even with a 100-fold reduction in the number of parameters. In particular, the AI can now easily detect partially obscured objects and complete their shape. This is the sixth US patent that Smart Engines scientists have received in 2023.

The new technology can be used to solve a wide range of problems: from document recognition to unmanned transport.

Often the vanishing point can be hidden behind other objects – billboards, other cars, lorries, etc.

The neural network architecture patented by Smart Engines scientists combines the blocks used in neural networks with the classic natural  image analysis tool, the Hough Transform. These two mechanisms complement each other. Hough Transform is often used to search and highlight straight lines, such as lines of text or the boundaries of objects – roads, houses, and documents. The convolutional layers of the neural network, in turn, help solve the problem of classifying the detected segments.

“The working mechanism looks like this. The first block of convolutional layers of the neural network builds local features of image points. The Hough Transform then integrates the local feature values along straight lines, making it possible to calculate complex nonlinear statistics such as variance. Convolutional layers after the Hough Transform work with these statistics along straight lines. The next stage is transposition, after which the feature maps transition to the original coordinates. At the last stage, the obtained result is processed by another block of ultra-precise layers, as a result of which we receive an image in the original coordinates, but at each point information from the entire image has been accumulated,” commented Alexander Sheshkus, Ph.D., the head of the machine learning department at Smart Engines.

Hough neural networks can yield much better results for the task of finding vanishing points, determining shapes, and highlighting the contours of objects, and are also better at detecting extended or partially obscured objects.

An important aspect of this invention is that the resulting neural network architectures are more resistant to a wide range of attacks on AI. Thus, even replacing part of the image will not become an obstacle to object detection. For example, if in an image with a road, the vanishing point is accidentally or intentionally obscured by some object, the Hough neural network will still find it.

If the edge of a document in the incoming image is blocked by a finger, artificial intelligence, knowing the shape of the document, can easily restore it.

Smart Engines patented this invention in the USA, thus receiving the sixth American patent for its inventions in 2023. Previously, the company received US patents for a number of its key inventions – a method for identifying documents, a method for combining frames in a video stream, a method for stopping text recognition in a video stream, and a method for using the Hough transform in networks. In addition, the company received a patent for a key invention in the field of computed tomography. Patented solutions are already used in Smart Engines software products for autonomous recognition of passports, ID cards, and other documents.