Uzbek traffic sign dataset for traffic sign detection and recognition systems

Abstract

Currently, deep learning algorithms are developing in machine learning and deep neural networks. However, deep learning algorithms still have pressing problems to solve. That is, the data set is insufficient and of low quality for machine learning. One of the ways to solve this problem is to data augmentation the dataset. There are two types of image data sets: labeled data sets and unlabeled data sets. These annotations are important for object recognition. Deep learning algorithms typically use an annotated dataset. Similarly, the You Only Look Once (YOLOv5) network model also requires an annotated dataset. In this article, we manually created annotations for each image using the labelImg software tool. Different countries have different Traffic signs. The dataset of traffic signs of Uzbekistan (UTSD) has not yet been developed. Therefore, in this work, we developed a UTSD dataset for use in a traffic sign detection and recognition system (TSDR). The UTSD dataset contains 3957 Traffic sign images belonging to 56 classes. We improved the UTSD dataset by data augmentation.