PNIPU: Synthetic Image Generator for Neural Network Training

Main article: Neural networks (neural networks)

2023: Making a Synthetic Picture Generator

Scientists of the Perm Polytechnic have improved the detection of objects by neural networks. The university announced this on March 1, 2023.

One of the tasks of neural networks is pattern recognition technologies. They are either individually or integrally used in areas such as security and surveillance, image scanning and creation, marketing and advertising, augmented reality, and image search. Training is a very important part of creating this technology. Too small or vice versa, a large amount of data in the neural network leads to incorrect operation. Sometimes even the optimal size of the data can lead to poor results if the objects on which the program was trained are captured from the same angle or are on the same background. As you know, specialists have to determine the boundaries of the studied objects manually in special programs. This process is very long and time consuming. Scientists of the Perm Polytechnic University have created a program with a generator of random synthetic images, which will allow training the neural network faster. The development is a special way to improve the quality of detection of the necessary objects, which allows you to ensure the technological sovereignty of Russia.

To facilitate the work of IT specialists, who usually manually create photographs to train the neural network, polytechnics have developed a program that generates synthetic pictures, combining images of a real object using a 3D camera, a naturalistic background and some noise effects - interference or environmental objects. For example, to create a set of pictures with a street lamp, the program additionally used tree branches that partially close the lamp, as well as rain, low light, camera defects. These noise effects make the resulting image more realistic. The quality of training depends on how evenly the data is mixed and how diverse the pictures are obtained.

Testing the program, we tried to conduct as many experiments as possible in order to get the largest possible overview of the impact of synthetic data on the performance of a neural network. The trial used data sets of 1000 and 2000 artificial pictures. After that, we noticed that such training gives poor recognition quality. For this reason, we decided to train the neural network, mixing synthetic data with real photographs, "said Leonid Mylnikov, associate professor of the Department of Microprocessor Automation.

Sets of images by topic based on synthetic data with a small number of real photos improved the quality of detection of an object by a neural network. This solves the problem of creating large databases necessary for training networks and greatly simplifies the work of specialists from IT the field. The technology can also be applied to moving images, "said Pavel Slivnitsin, graduate student of the Department of Information Technologies and Automated Systems.

According to the developers, the program they created is a universal tool that can be used in any area. According to them, the concept of creating artificial data sets is suitable for any type of neural networks, which is especially important, since each is distinguished by its own technological features and established fields of application.

As of March 2023, polytechnics are engaged in obtaining even more realistic images, for example, containing elements such as the effects of corrosion and deformation of an object studied by a neural network. This will further improve the quality of detection.