Probing into the Robustness of Deep Learning Models in Visual Recognition Applications