| Facial expression recognition technology refers to a set of face images that have been subdivided artificially and labeled with a facial expression category label,uses machine learning method to establish a facial expression classification and recognition mode,and then uses the established classification mode to realize the understanding of human facial expressions and the judgment of psychological activities.It is the basis of intelligent human-computer interaction.And it is widely applied in many fields.The key factor affecting the accuracy of facial expression recognition is machine learning algorithms.With the appearance of deep learning algorithms that have ability more abstract high-level attribute classes or feature representation capabilities through the combination of low-level features,the deep learning model represented by supervised learning model:Convolutional Neural Network(CNN)and Non-supervised learning model:Deep Belief Network(DBN)is favored by many pattern recognition researchers.The advantage of deep learning algorithms are that even if face detection,feature extraction and other preprocessing are not performed,as long as the face image are large enough,relying on its own excellent distributed feature representation and memory capabilities,the accuracy of facial expression recognition based on it is obviously higher than other traditional methods.However,in practical applications,limited by computer hardware conditions,the process of determining the structure of deep learning network and its related parameters by experimental experience is very time-consuming(even the system runs out),which makes the researchers miserable.Therefore,the research of face detection and local feature extraction method is carried out to reduce the data dimension of the input side of deep learning model,which is conducive to improving the efficiency of learning and recognition.At the same time,facial expression is a kind of non-rigid detailed texture feature.By extracting the local detailed texture features of the face,facial expressions can be used for deep model learning and recognition,which undoubtedly helps to further improve the accuracy and robustness of expression recognition.In view of this,the main research contents of the paper include:(1)Facial expression recognition technology was established based on the combination of face detection,local feature extraction and deep learning algorithms,and the AdaBoost face detection method based on haar-like characteristics was realized.When the hardware conditions of the deep learning platform were not improvedsignificantly,and the numerous experimental processes were frustrated,through theoretical analysis,the technical route of face detection and local feature extraction for face images is established before the deep learning algorithms are trained in the network.Considering the detection rate and accuracy index,the advantages and disadvantages of various face detection algorithms were compared and analyzed,and the AdaBoost face detection method based on haar-like characteristics was studied and implemented.The comparison experimental results show that the AdaBoost face detection method based on haar-like feature has better detection efficiency than the face detection method based on skin color model.(2)An improved Active Shape Model(ASM)local feature extraction method was proposed.A fatigue expression recognition rule was established and an application system for driver facial fatigue state recognition was implemented.On the basis of the implemented face detection method,faced with fatigue face recognition,through the deep research of ASM theory,on the basis of the Helen database,ASM local feature extraction method based on 194 face feature points was first realized at home and abroad.The use of PERCLOS parameters to determine eye fatigue status was combined with a Support Vector Machine(SVM)classifier to determine the degree of head sagging,a fatigue expression recognition rule was established,and a real-time application system for driver facial fatigue status recognition was developed.(3)An improved facial expression recognition method based on the combination of Local Binary Pattern(LBP)feature extraction and DBN was proposed.Aiming at seven kinds of emotional expression recognition application of CK+database,which are neutral,happy,anger,sadness,surprise,fear and disgust,a neighborhood smoothing operator is established,and an improved LBP feature extraction method is proposed.On this basis,DBN network structure and parameters were determined and the face expression recognition method based on DBN was realized by a large number of experiments.The experimental results show that before learning CK+ database in DBN network,using the AdaBoost algorithm based on haar-like feature for face detection and the improved LBP model to extract the local features of human face were beneficial to shorten the time of DBN network training,and at the same time,the accuracy of expression recognition was improved obviously.(4)A real-time application system of expression recognition based on CNN depth model is developed.Using the lightweight CNN deep model and training for CK+ database,thehuman-computer interaction platform was built.An application system that meets the need for real-time recognition of seven emotion expressions such as neutral,happy,anger,sadness,surprise,fear and disgust.The speed of human-computer interaction system is fast,and the average expression recognition accuracy reaches 84.4%. |
