Research On Facial Expression Recognition Method Based On Key Expression Area

Over the years,with the rise of artificial intelligence,people more and more hope that machines can provide more convenient and efficient customized services more intelligently.An important function of intelligent machines is intelligent human-computer interaction.Therefore,the demand for intelligent human-computer interaction lies in Facial expression recognition is an important way to realize intelligent human-computer interaction,and convolutional neural networks have extremely strong performance in image processing.In recent years,facial expression recognition(Facial Expression Recognition,FER)based on CNN The research has been uninterrupted,and many results have been achieved,which greatly promoted the development of facial expression recognition research.At present,the facial expression recognition method based on neural network has some limitations,which has led to the limitation of the practical application of previous research.For example,the existing facial expression recognition method is basically based on the positive facial expression recognition under the experimental control environment.In an open environment,or in a native environment with relatively large changes,its generalization performance is poor,and it may even become unusable.Specifically,the limitations of current facial expression recognition methods based on neural networks are mainly manifested in three aspects:(1)Without considering the strength of facial expression information and the relationship between different regions of the face,the method of facial recognition is directly used The facial expressions are extracted from the entire face and used for facial expression recognition.The facial expression features are concentrated in the areas of the eyes and eyebrows,mouth,and lower part of the nose.These areas are called key expression areas,which are more valuable for expression recognition.(2)Some related work considers the key expression area,but the extraction and use of the key expression area is divided into two independent stages,which makes the performance of the overall model uncontrollable.(3)The current expression recognition methodusing key expression areas has certain requirements for facial pictures,such as [8][22],etc.When the requirements are not met,the model cannot be used.In order to break through the limitations of previous facial expression recognition research,this paper based on the key expression area for facial expression recognition,proposed a solution,and carried out experimental verification.The main research contents of this article are as follows:1.In view of the first two limitations,this paper analyzes the correlation between facial expression classification accuracy and the boundary of key expression areas,and proposes a hybrid pre-training method for expression recognition and boundary box regression of key areas;analysis of similar facial expression pictures and facial expression recognition Based on the relationship between the probability distribution of facial expressions predicted by the model and the performance of the model,a fine-tuning method for the model based on narrowing the output difference within the class is proposed;finally,the two are integrated to form a two-stage training method for training the facial recognition model.2.Based on the research of the two-stage training method,this paper also analyzes the three limitations of the previous expression recognition research,and analyzes the two-stage training method of the facial expression recognition model proposed in this paper and the previous fusion expression recognition method.On the premise of facial expression recognition in the region,a key expression area discrimination model is further proposed to generate a mask image of the key expression area to reduce the quality of facial expression pictures.A dual-channel feature fusion model is proposed to improve expression feature extraction Ability to introduce a concept of key expression area coincidence and integrate the training process of these two models to form an end-to-end dual-channel facial expression recognition model based on feature fusion.3.The two-stage training method for expression recognition model training and the two-channel facial expression recognition model based on feature fusion proposed in this paper are respectively tested.The recognition accuracy of the model trained by the two-stage training algorithm on CK + reaches 93.59%.The recognition accuracy of the dual-channel facial expression recognition model based on feature fusion onCK + and FER2013 reached 94.63% and 72.59% respectively,which basically reach the results obtained by the current better expression recognition methods.