Loading...

Face Recognition Based on a Single Training Image for Each Person Across Large Pose Variations

Imanpour, Nasrin | 2013

750 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 45198 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Manzuri-Shalmani, Mohammad Taghi
  7. Abstract:
  8. Most existing face recognition methods require several gallery (training) images per person for optimally training the system. On the other hand, in many real applications of face recognition such as law enforcement, driver’s license and identification by passport, generally only one image per person is available. However, “the variations between face images with different illumination and gestures are almost always larger than image variations due to change in face identity”, Therefore, in this thesis we have considered face recognition with changes in ambient lighting conditions and viewing angles where only one training image per person is availabe. A method have been proposed in which we determined the position of the eyes and then we extracted half of the face using some number of extracted patterns and then we tried training each part of the face such as the mouth, eyes and the whole face using maximum likelihood. Then we determined recognition rate of each part of the face and also the whole face. By means of this consequently we applied these weights to the final face recognition system. We have used color FERET and CMU-PIE databases for traing and testing purpose. In this study, we determined that features such as eyes and mouth are more resistant to rotation. And using these features improved the recognition rate about one percent, especially for the angles of rotation higher than 22 degrees. We get the recognition rate of 100% for faces with no angle and near 90% for faces with few angles of about 20 or less. For angles more than this our reconition rate decreased. For example on CMU-PIE database with angle 62 degree the recognition rate was 61.98%. There is no feature based mapping to that can get better results
  9. Keywords:
  10. Face Detection ; Detection ; Gabor Wavelet Transform ; Single Training Image Per Person

 Digital Object List

 Bookmark

No TOC