| Summary: | 碩士 === 國立政治大學 === 應用數學系 === 106 === Since the need of better security, more and more face recognition related research papers have been given in recent years. Their results are widely used in various fields, such as smart car (or self-driving car), FinTech, smart retail, robot, drone, business analysis, and crime prevention. However, when the content of images, such as head posture, lighting, complex background, and aging, has a big change, it is harder to recognize the right person. Therefore, the question of factors that influence the recognition result and how to improve the system recognition rate becomes an important research topic.
This paper first compares several common dimension reduction and classification techniques of multivariate analysis methods, including principal components analysis, linear discriminant analysis, two-dimensional principal components analysis and two-dimensional linear discriminant analysis, for feature extraction. We divide the data in each of our four databases into two halves. The first half is for training, while the second one is for testing. The empirical results show that when the changes of head postures are small, the two-dimensional linear discriminant analysis has a very good correct classification rate, which is 94% on average. The linear discriminant analysis has the second highest correct classification rate, which is 92% on average. In addition, if we pre-process the images, the correct classification rate increases a lot on each of principal components analysis and two-dimensional principal components analysis.
Finally, we give a new updating formula for computing covariance matrix. Using this new updating formula and our face recognition technique of principal components analysis. We develop a Graphical User Interface, which can unlock any personal computer. When new face image information is given, we update the covariance matrix through our proposed iteration method, which can easily keep the data for the face recognition in the latest and the most complete state without recalculating the huge and complicated covariance matrix.
|
|---|
