Пролеев Лев Николаевич
One-Shot Learning for Traffic Signs Recognition
Прикладная математика и информатика
Traffic signs recognition is a problem of understanding the class of the traffic sign on a given image. In the recent years, there was a massive improvement in quality of deep machine learning models. However, such models require a large labeled dataset to train on. Traditional deep learning models trained on a traffic signs dataset for one country generally cannot be used on a set of another country’s traffic signs without additional training on corresponding data. Acquiring a new dataset of appropriate size is still a challenging problem. One-shot learning framework though implies effective usage of models after training on only a few labeled examples. Therefore, high-quality models of such kind can lower the burden of data labeling for people and help to adapt trained models more easily. This paper gives an overview of existing one-shot learning methods and analyzes several approaches for training feature extractor for traffic signs classification, such that it will be representative even on image classes that were not present during training.
Текст работы (работа добавлена 21 мая 2018г.)