الفهرس 
Face Recognition SystemsOnly 14 pages are availabe for public view
 |  | from | 230 |  |  |
| | | from | 230 | | |
Abstract
The thesis presents a comparative study of performance for face recognition algorithms in order to select the algorithms that have the highest performance and overcome the problems faced in recognition due to plastic surgery. A plastic surgery database that contains face images with different types of surgeries is used. The work reports the performance evaluation of eleven photometric illumination techniques, five histogram normalization techniques and four feature extraction techniques. A minimum distance classifier has been adopted and four distance similarity measures were used. Face identification/verification techniques are considered in the present work. Experimental results showed that for identification purpose and for all types of plastic surgery, best illumination technique is Gradient faces (GRF) normalization Technique, no histogram normalization and the best feature extraction technique is gabor principal component analysis (GPCA). Mahalanobis cos (MAHCOS) distance gives the best results. This results in a correct recognition rate of 96.55%, 85.11%, 84.91%, 81.25%, 78.08% and 73.20% in case of resurfacing, Blepharoplasty, Forehead-lift, Fat-injections, Otoplasty and Rhytidectomy plastic surgery, respectively.
For the verification problem and for all types of plastic surgery, best illumination technique is GRF Normalization Technique, the best histogram normalization is Histruncate histogram and the best feature extraction technique is gabor kernel fisher analysis (GKFA) using MAHCOS distance. The minimum error rates reach 0.0140, 0.0173, 0.0451, 0.0463, 0.0549 and 0.0667 in case of resurfacing, Bleharoplasty, forhead-lift, Otoplasty, Rhytidectomy and fat-injections, respectively.
For both face identification/verification the minimum distance classifier using MAHCOS distance gives the best results.
It is concluded that the face identification/verification techniques adopted in the present work are promising and could be used and developed to improve the performance of plastic surgery face recognition systems and to find the nearness between the pre-plastic surgical face to the-post plastic surgical face.