Research On Face Expression Recognition Based On Video

In recent years,with the rapid development of fingerprint recognition,face recognition and other technologies,the communication between people and computers becomes more and more frequent.In the process,computer is required to understand human emotional states and recognize facial expressions.Facial expression recognition has become one of the current research hotspots.It has board application prospects and important research value in online education,game entertainment,medical care,fatigue driving and other aspects.Up to now,static image and video sequence are mainly research objects of facial expression recognition.Since the expression is a process of change,the static image is difficult to reflect the dynamic characteristics of the expression,which makes it difficult to improve the recognition rate.The video sequence can provide motion information of expression changes.Therefore,this thesis takes video sequence as the research object,which has greater research value and practical significance.This thesis mainly studies the following aspects.1)Detect and locate the face in the video sequence.The method used in this thesis is to combine skin segmentation algorithm with the AdaBoost algorithm.Firstly,the skin color segmentation algorithm is used to segment the skin color regions as candidate regions,which can eliminate the interference of non-skin color regions such as background.Meanwhile,AdaBoost classifier based on Haar-like features is trained to construct cascade classifier.Then,the candidate regions are used as the input of the cascade classifier,and finally,the output results are obtained from the cascade classifier.2)Feature extraction is performed on the face in the video sequence.Considering the dynamic texture feature and dynamic geometric feature of the face,the LBP-TOP algorithm is used to extract the dynamic texture feature,and the motion coding algorithm is used to extract the dynamic geometric feature.By combining the two feature in series,a video sequence can be represented by a feature vector with the dimension of 200.3)Segment the long video sequence containing multiple expressions.Different from the video sequence in a public database,in an actual application,it is usually necessary to identify multiple expressions in a long video,so it is very important to accurately segment the video sequence.In this thesis,the similarity between the input sequence and the reference frame is calculated by defining the neutral expression as the reference frame.According to the variation law of similarity,the starting frame of the expression is obtained by using the discriminant formula.Finally,video sequence is segmented according to the starting frame.4)The expression classification is performed according to the feature vector of the video sequence.Due to the advantage of SVM algorithm in the classification of small samples and non-linear data,this thesis uses support vector machine and structured support vector machine as the binary classifier,and then combines multiple binary classifiers into multiple ones classifier by the method of one-to-many combination.Finally,experiments were carried out on the facial expression database CK+,and the dynamic texture features,dynamic geometric features and mixed features of the visual sequence were extracted to identify the effectiveness of the algorithm.