| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 101 === Recently, there is a significant progress in study of face recognition. Meanwhile, hardware technology advances, too. The face recognition has been applied on more areas. In order to make the face recognition practical, it is required to reduce the effect of variations on performance. In this thesis, we focus on the pose variations problem and take Tied Factor Analysis as the core of system. The concept is to explain the complex distribution caused by pose variations with several Factor Analysis models. Moreover, there exists a certain representation for images that at different poses from the same subject without regard to pose and facial variations. In learning process, the EM algorithm is used to find the optimal parameters iteratively. In recognition process, a variety of generative models are designed to explain the different generative procedures of data, and then take the most likely procedure as the result. To increase the tolerance for variations, we use the Affine Transform normalization to reduce the effect of pose variations. And we add various variations to promote the performance.
|
|---|
