Development of a Method for Generating Images of Road Signs for Training Neural Network Models to Recognize Images in Video Streams

Abstract

The paper describes a mechanism for generating images containing road signs for training neural networks in pattern recognition. Faster R- CNN convolutional network was used as a neural network model. The proposed image generation method takes into account the specifics of the subject area and simulates possible distortions that make recognition difficult. During the experiments, an increase in recognition efficiency was observed as a result of adding generated images to the training set. Since generating one image requires significantly less resources than obtaining a real image, especially when achieving the scaling effect, the number of synthetic images during the experiment exceeded the number of real images used by orders of magnitude.