Facial Expression Recognition Based On Key Parts Of Deformation

As an effective non-verbal means of normal interaction,facial expression is the basic way to express our feelings.Dynamic facial expression helps communicators to understand each other's words accurately,which make the study of facial expression intelligent recognition meaningful.By consulting domestic and foreign materials related to facial expression recognition,further studying available methods and theories related to recognition and aiming at common procedures of facial expression recognition,I introduce every step from face detection to facial expression recognition in detail.Effective multi-angle face detection can be achieved based on mixed model of face.Then,the posteriori probability Snake model can help with the specific location of facial features and the division of parts which have great influence on facial expression recognition.Finally,by retrieving the PHOG characteristics of key deformation positions and applying sparse coding way,facial expression recognition can be achieved.What's more,noise test and shelter test have proved that its' robustness is good.The thesis is briefed as below:1.In view of the actual situation in which face images collected by cameras are not always regular,I mainly focus on the study of multi-angle face detection.Based on the nonrigidity and variability of face,I choose HOG feature which can better describe the marginal information of target.In order to better detect multi-angle face,build mixed model which consists of the local deformation model of front face,half of side face and front of side face,and realize the multi-angle face detection,I use LSVM classifier to train the local deformation model of multi-angel face.2.Almost all the intensive study of face detection is based on the extraction of facial features,which requires the accurate alignment and segmentation of facial features.The thesis conducts a positioning study on facial features by means of Snake model matching and improves the minimizing process of deformation model energy by probability distribution to calculate the maximum posterior probability,which can overcome Snake model's defects of being sensitive to initial position,solve the robustness of the location of facial features effectively and improve the positioning speed.It can also divide the deformation areas which is distinctive in the expression changes,and that is beneficial to the future facial expression recognition.3.The thesis focuses on the study of the recognition of static facial expressions.Expression recognition is conducted by sparse coding which is proposed by stimulating the perception of neuron.In order to get the expression dictionary and prepare for the recognition of sparse coding,we need to train the PHOG characteristics of critical area by means of ODL.Finally,based on the simulation experiment on noise and shading expressions by means of PHOG+SVM,I analyze the results and get that the thesis is of great robustness and is effective in facial expression recognition.