Researches On Facial Expression Recognition Based On Deep Feature Learning Algorithm

Facial expression,an essential way for information transmission in interpersonal communication,plays a very important role in detecting the speaker's feelings,comprehending the meaning of speech,and capturing emotional details.With the rapid development of information technology and computer science,people have put forward higher requirements for the “intelligence” of machines,and facial expressions are the fastest and most effective way for machines to understand people.Therefore,The demand for automatic facial expression analysis technology in the fields of robotics,medical treatment,driving,etc.is becoming more and more extensive.So how to improve the performance of facial expression recognition(FER),enhance the model's robustness to various external disturbances,and adaptability to environmental changes are important problems that need to be solved urgently.This paper combines the current technical difficulties and research hotspots in the field of FER,and makes the following two main contributions in deep features learning:(1)A more discriminative feature learning method is proposed in this paper.Under nonexperimental conditions,facial expression recognition systems will be affected largely by the variations of the backgrounds,face poses,lighting,occlusion,individual differences and so on.This makes the features learned by deep neural network coupled together in the feature space,unable to be distinguished between the classes.To solve this problem,we developed a more powerful deep feature learning method——Fisher loss,under the inspiration of the Fisher separability criterion.Minimizing this loss is equivalent to minimizing intra-class dispersion and maximizing inter-class distance in feature space,where the deeply learned features can show better separability.A series of experiments were conducted on MNIST and FER2013 data sets.The experiment and visualization results show that Fisher loss has advantages over the softmax,center loss,as well as the island loss,and we got competitive performances on these datasets in our experiments.(2)A local correlation attention mechanism is proposed to learn cross-region correlation features of human faces.Under the complicated and changing shooting conditions of the machine,it is very difficult to capture the key information of facial expressions.CNN focuses on the local receptive field of the image.It needs to stack more convolution modules to capture the long-distance dependencies between different areas of the face,which is a very inefficient way.To solve these problems,this paper proposes a local correlation attention mechanism,which can learn the correlation information from different part of the face,and fuse the features across regions,making the model accurately focus on the key areas related to expression,largely improving the accuracy on facial expression recognition.In the experiment,we compared the difference between the proposed model and the baseline model.It is found that the local correlation attention mechanism can bring an average accuracy improvement of 1.7% in the three data sets of FERPlus,SFEW,and Affect Net.Compared with the methods of the last two years,we have achieved State-of-the-art results on the SFEW data set to our knowledge,while the performance on FERPlus and Affect Net can be comparable to others.Besides,Grad-CAM is used for heat-map visualization.Compared with the baseline model,we found that the local correlation attention mechanism has its unique advantages in three aspects: Efficiently learning cross-region correlation features,Accurately locating expression region and Effectively shielding image interference information.