Improving performance of the PRYSTINE traffic sign classification by using a perturbation-based explainability approach