My image classification model badly performs with complex images, how to resolve?

Question

I have a traffic sign dataset, after training it with a model i created the val_accuracy is around 93% which is good however I have a mixed behavior when I try to predict a