Assessing the Impact of Various Image Corruptions on the Recognition of Traffic Signs by Machine Learning
Chuwen Song · 2020
Evaluating the traffic conditions correctly is a critical task for safe autonomous driving. Recognizing the traffic signs flawlessly is one big component. It may seem like a simple task given that there are a limited number of traffic signs available. However, the traffic sign images may become corrupted when the weather condition is bad, there is some blockage from the tree or when the traffic sign itself was damaged. This study, by evaluating the robustness of traffic sign recognition against these scenarios, provides valuable insights into image corruption and corruption of datasets in general for machine learning. It raises potential concern for applying a well-performed machine learning algorithm to an unseen new test data in a new environment, and would emphasize the importance in training the model on enhanced training sets with various types of preassigned image corruption.