Traffic sign classification using cost effective data augmentation
René Pihlak, Andri Riid, Sadok Ben Yahia · Procedia Computer Science · 2025
In recent years, the development of computer vision and deep learning in general has significantly impacted various domains. As (semi-) autonomous driving and advanced driving assistance systems gain popularity, the demand for robust and accurate traffic sign classification solutions continues to rise. The task of classifying traffic signs faces a special challenge in the European Union, where cars often cross national borders and traffic signs differ from country to country. We propose a method to augment a training dataset with illustrative drawings to improve the precision of traffic sign classification models. Our results suggest that illustrative drawings improve precision. For example, the precision of Estonian traffic sign classification models increased from 0.92000 to 0.92667 when using illustrative images of Estonian traffic signs, and even using illustrative images of German traffic signs resulted in an improvement (0.92222). Similarly, the precision of German traffic sign classification models increased from 0.78889 to 0.83111 with German illustrations and 0.83556 with Estonian illustrations. Moreover, the results suggest that models trained on optimal ratio of real traffic signs and illustrative images also improve F1, recall and precision of predictions.