DATASETS PREPARATION FEATURES FOR TRAINING ARTIFICIAL NEURAL NETWORKS

Abstract

The article analyzes key aspects of the formation of training datasets, including images of objects, for training artificial neural networks, in particular on the YOLO platform, which is widely used for recognition tasks. The subject of the study is the process of identifying and classifying objects associated with track scenarios, such as track robots, barriers and track surfaces. The article presents an approach to creating two-dimensional images with the appropriate annotation of the classes "robots", "track" and "barriers". Particular attention is paid to data augmentation methods that allow increasing the variability and improving the quality of training samples, which is critically important for the effectiveness of training neural networks. The study describes the process of specializing a neural network in recognizing objects by specific characteristics or in conditions with limited resources. After training, the network is used to identify and segment certain objects in images. The collected recognition results are subject to aggregation, which allows integrating information from different sources, thereby increasing the overall accuracy of the system. The proposed methods can serve as the basis for developing and improving object recognition algorithms in various applied tasks, such as mobile robotics, monitoring and environmental analysis. The relevance of the work is due to the need to increase the stability and accuracy of object recognition systems in difficult conditions, in particular in the presence of limited and low-quality training data. The obtained results and approaches can be applied to solve practical problems and support further research in this area.