Research On Facial Expression Recognition Based On Manifold Learning

As the most direct and basic ways of communication for human emotions,expression is a very effective way in non-verbal communication. People not only canexpress their thoughts and feelings accurately and subtly through a variety of facialexpressions, meanwhile, it can identify the status of others and the inner world of eachother’s face. Therefore, the expression has also become one of the most abundant bodylanguage to participate in various activities of the human. If the computer andintelligent robots have emotional understanding and expression like humans, this willimprove the level of human-computer interaction fundamentally to a higher level so asto enable the computer to have a better serve for our human.With the depth improved of face detection, tracking and face recognitiontechnology in recent years, facial expression recognition research has graduallybecome a hot research field of pattern recognition and artificial intelligence. Currently,face recognition technology is still at the research stage, there are many problems needto be solved and to have further in-depth study. Such as how to extract facialexpression feature accurately, how to reduce the redundant information of theexpression feature of high-dimensional maximum and estimate expression intensityand their classification accurately, and how to conduct multi-disciplinary integration tohave multi-point analysis for expressions.This paper studies the research status and some application problems of facialexpression recognition at home and abroad systematically, and proposes or improvesome algorithms for some key issues, the main research work includes:(1) In the pretreatment of human facial expression, for accessing and positioningof the expression, we propose the acceleration algorithm of the cascade classifier basedon the dual-threshold for face detection, and apply the AdaBoost algorithm to thehuman eye positioning. It greatly improves the classification accuracy of expression inthe latter time, and lays a good foundation for the stage of pre-processing ofexpression.(2) As basis of contribution of the key features of facial expression, extracts themost critical Gabor wavelet-sensitive feature information for expression recognition, itcan reduces post-processing redundancy and selects the most informative expressioncharacteristics to improve the discrimination of various expression.(3) Combine the above key expression feature, the secondary embedded manifold facial expression recognition algorithm which is based on local sensitive informationand Laplace Maximize Discriminant Analysis(LMDA) is proposed. Using PCA to haveinitial dimensionality reduction for local sensitive information and has a adaptiveweighted for between-class features which are got by initial dimensionality reduction,On this basis, we establish laplace within-class and between-class scatter matrix whichis based on the manifold, and maximize the matrix to obtain the projection matrix.Lastly, using certain distance criteria to have expression classification validation on thefacial expression recognition libraries. The experimental results ultimately prove thatthe algorithm has a good recognition effect and strong robustness.