A Method For Face Recognition Based On The Alignment Of Facial Reference Points

The technology of face recognition plays an important role in computer vision andmachine learning. It can be widely used in many areas such as information security,human-computer interaction, image retrieval and so on. Face recognition is identifying theface detected and extracted from static image or video. Its difficulty is that the faceappearance is susceptible to the influence of the factors such as posture, shielding,illumination, expression and time delay. The current researches separate feature pointslocating and feature matching while the work on the collaborative ability between them isvery scarce.In this thesis, the current feature matching algorithms have been roughly divided intothree categories, face recognition based on geometric characteristics, face recognitionbased on statistical characteristics and others. The first one trains vectors constructed bygeometric feature and gets recognition classifier. It is simple but difficult to extract stablefeatures. The second one projects the faces into a new linear space, then establishes adecision boundary so as to achieve the effect of classification, but the scope of applicationis restricted. The situation of the third one depends on its model.There are two types of feature points locating algorithm，one based on optimal and theother based on regression. The locating effect of the former depends on the design andoptimization of error equations, however, that of the latter depends on the parametermodel. This thesis described a face alignment algorithm based on double regressionwithout parameter model in detail, analyzed its double-layer structure, working principleand training feature and improved the second layer by random forests. Experiments showthat the improved model is better than the original.In addition, this thesis explored the collaborative ability between the feature pointslocating and feature matching. Firstly, introduce the face recognition algorithm based onlinear discriminant analysis and its improved version. Secondly, combine the improvedface recognition algorithm with the improved face alignment algorithm. Thirdly, extractthe Gabor feature and geometric feature. Experiments show that the combination offeature points locating and face recognition can improve the accuracy rate of recognition.