Research On Face Detection And Expression Recognition Algorithm In Rotation-invariant

Facial expressions contain a lot of information,and its recognition technology gives computers the ability to perceive human emotions.It has a large number of applications in the fields of education,medical treatment,safety,driver fatigue monitoring,robotics,etc.,and has important academic and application value.In automatic facial expression recognition,pose change is one of the most important obstacles in automatic facial expression recognition.It can significantly change facial features,which greatly affects the accuracy of recognition tasks.This thesis proposes solutions to the pose problem in video facial expression recognition,combined with convolutional neural networks,from face detection and expression feature extraction,respectively.The main research contents of the thesis are as follows:(1)Face detection is the basis of relevant recognition tasks such as facial expressions.Focusing on the problem of how to detect faces with pose changes in the video stream,this thesis combines facial landmark with pyramid optical flow,and proposes a rotation-invariant face detection algorithm based on cascade network and pyramid optical flow.First,the cascaded progressive convolutional neural network is used to locate the face position and facial landmark of the previous frame in the video stream;secondly,to obtain the optical flow mapping between the facial landmark and the face candidate frame,use an independent facial landmark detection network to detect the current Reposition the frame;then calculate the optical flow displacement of the facial landmark between the two frames before and after;finally,through the mapping relationship between the optical flow displacement of the facial landmark and the face candidate frame,correct the face detected in the video to complete the in-plane rotation Face invariance detection.By comparing the two public data sets with related detection algorithms,the experiment shows that the algorithm has high accuracy and can effectively solve the problem of rotating face detection in the plane.At the same time,the detection speed of the algorithm in this thesis has a great advantage,and the problem of the jitter of the face candidate frame in the video stream has been well solved.(2)The effective extraction of facial expression features is the core step of facial expression recognition.Focusing on how to extract facial features reliably and efficiently,this thesis uses Mobile Net V2 as the basic feature extraction network,and associates key facial features with global features through attention mechanisms,and proposes facial expressions based on lightweight networks and attention modules Identification algorithm.First,get the face image in the video stream,and use the Mobile Net V2 network to get the features of the facial landmark area;second,use the attention module to assign attention weight to each area feature,and aggregate the different area features generated by the network into a global Feature;then,by associating the facial landmark region feature and the global feature of the face,a joint feature is obtained;finally,the recognition result is output according to the joint feature.At the same time,in order to improve the training performance of the facial expression recognition model,on the basis of the original data set,organize and label the corresponding pose data.After comparing with the related expression recognition algorithms in the three public data sets,experiments show that the algorithm architecture greatly improves the extraction of facial expression features,and the algorithm recognition rate is better than related advanced algorithms.