Method for creating synthetic data sets for training neural network models for object recognition

Introduction: The lack of training data leads to low accuracy of visual pattern recognition. One way to solve this problem is to use real data in combination with synthetic data. Purpose: To improve the performance of pattern recognition systems in computer vision by mixing real and synthetic data for training, and to reduce the time needed for preparing training data. Results: We have built an intelligent information system on the basis of the proposed method which allows the generation of synthetic images. The system allows to generate large and representative samples of images for pattern recognition neural network training. We have also developed software for the synthetic image generator for neural network training. The generator has a modular architecture, which makes it easy to modify, remove or add individual stages to the synthetic image generation pipeline. One can adjust individual parameters (like lighting or blurring) for generated images. The experiment was aimed to compare the accuracy of pattern recognition for a neural network trained on different training samples. The combination of real and synthetic data in model training showed the best recognition performance. Artificially generated training samples, in which the scale of background objects is approximately equal to the scale of the object of interest, and the number of objects of interest in the frame is higher, turned out to be more efficient than other artificially constructed training samples. Changing focal length of the camera in the synthetic image generation scene had no effect on the recognition performance. Practical relevance: The proposed image generation method allows to create a large set of artificially constructed data for training neural networks in pattern recognition in less time than it would take to create the same set of real data.