The Study And Implementation Of Facial Expression Recognition Algorithms Based On Deep Learning In Natural Environment

In view of the theoretical value or application value of facial expression recognition(FER)technology in many fields such as intelligent human-computer interaction(HCI),market,education and psychology,it has received more and more attention in academic circles and industry in recent decades.Although the technology has made more progress,when it is applied to the natural environment with less constraint or no constraint,it is still susceptible to factors such as head posture,ambient illumination,identity difference and registration error.Related researches have revealed that we can effectively alleviate above effects by improving the model's ability to represent expression images.Although convolutional neural network(CNN)has strong representation ability for images,it is still insufficient for facial expression recognition.The research goal in this paper is to explore and design some models based on the CNN with more powerful expression representation ability,in order to identify seven predefined basic facial expressions more accurately and robust from face images in the natural environment.The main research content of this paper includes the following two parts:(1)Most of the convolutional neural networks for facial expression recognition use the classic maximum pooling and average pooling.These pooling strategies limit the model's representation ability by retaining only the first-order statistical information of the input features,so a variety of second-order pooled convolutional neural network structures are designed in conjunction with baseline convolutional neural network and second-order pooling to capture second-order statistical information of input features.(2)Most convolutional neural networks for facial expression recognition contain only feed-forward connections.The human brain,however,contains far more feedback connections than feed-forward connections and exhibits strong visual attention ability through feedback mechanisms.We introduce a feedback layer in the above-mentioned baseline convolutional neural network to construct feedback mechanism to obtain a feedback convolutional neural network,in an attempt tosimulate human visual attention ability,and further to capture the expressions related information distributed in various local areas of the face more efficiently.All the models involved in this paper are tested in the same way on the RAF-DB expression dataset.The experimental results show that the recognition performance is better than the baseline model whether it is a second-order pooled convolutional neural network or a feedback convolutional neural network.