Impact of Image Corruptions on the Reliability of Traffic Sign Recognition Using Machine Learning Technique

Abstract

As autonomous driving gets closer to be widely applied, it is important to guarantee that traffic signs are recognized correctly. Due to change of light conditions and blockage by some other objects, traffic signs can sometimes be partially corrupted. In this paper, we evaluated how machine learning would respond to different types of image corruption by various degrees. Removing a higher percentage of pixels will gradually harm the recognition accuracy, and removal by blocks causes the biggest harm, which is consistent with human observation. Changing to various colors or a single color doesn t seem to cause significant differences. This study, by building a model for traffic sign recognition and evaluating its robustness to various types of image corruptions, provides insights into corruptions of datasets for machine learning in general, and provide potential concern for applying the well-trained model to a new test set with corruptions, which could be a concern for applying autonomous driving in certain areas.