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

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

The Smart Engines company received an American patent for proprietary development. With its help, AI can restore the parameters of three-dimensional space from a two-dimensional picture. Scientists proposed to use the Hough transformation as layers of a neural network, thanks to which neural networks – with a 100-fold reduction in the number of training parameters – cope much better with the basic tasks of computer vision. In particular, now AI can easily detect partially obscured objects and complete their shape. It sixth US patent received by Smart Engines scientists during the year.

The new technology can be used to solve a wide range of tasks – from document recognition systems to unmanned transport systems.

The neural network architecture patented by Smart Engines scientists combines the building blocks used in neural networks with a classic real-world image analysis tool, the Hough transform. These two mechanisms complement each other. Huff analysis is often used to find and select straight lines, such as lines of text or boundaries of objects – roads, buildings, documents. And the convolutional layers of the neural network, in turn, help solve the problem of classifying the detected segments.

Illustration of the work of Hafiv neural networks

The working mechanism looks like this. The first block of convolutional layers of the neural network builds local features of the image points. Then the Hough transform integrates the values ​​of local features along straight lines, as a result, it becomes possible to calculate complex nonlinear statistics – for example, variance. Convolutional layers after the Hough transform work with these statistics along straight lines. The next stage is transposition, after which the feature maps are transferred to the original coordinates. At the last stage, the obtained result is processed by another block of convolutional layers, as a result of which we receive an image in the original coordinates, but at each of its points information from the entire image is accumulated, – comments the head of the Smart Engines machine learning department, Ph.D. Oleksandr Sheshkus.

Haffian neural networks can better cope with the task of finding points of origin, determining the shapes and highlighting the contours of objects, and also better detect 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 AI attacks. Yes, even replacing a part of the image will become an obstacle to detecting the object. For example, if the sunrise point is accidentally or intentionally obscured by some object in the image from the road, the Hafi neural network will still find it.

Often, the starting point can be blocked by extraneous objects – billboards, other cars, trucks, etc.

If the edge of the document in the input image is covered by a finger, artificial intelligence, knowing the shape of the document, will be able to easily restore it.

Smart Engines patented this invention in the United States, thus receiving the sixth American patent for its developments in 2023. It previously received US patents for a number of its key inventions – method document identification, method integration of frames in a video stream, a method of stopping text recognition in a video stream and method using the Hough transform in networks. The company received another patent for a key invention in the field of computer tomography. Patented solutions are already used in Smart Engines software products for autonomous recognition of passports, ID cards and other documents.

Related posts