By building two models including Deep Convolutional Neural Networks and Deep Sparse Rectifier Neural Networks on facial expression dataset

we made contrastive evaluations in facial expression recognition system with deep neural networks.Based on prior structure knowledge of facial expression

we proposed a fast and simple improved method called K Region Of InterestK-ROI

which relieved the poor generalization of deep neural networks on experimental dataset due to insufficient data and decreased the testing error rate apparently and generally.Finally

we infer the experimental results and analyze comprehensively for the possible validity with deep neural networks on arbitrary image dataset.