Comparison of face recognition techniques: the effects of features and parameters setting
Face recognition is an important biometric application. In this thesis, a comparison of two common techniques for face recognition is carried out under the same conditions. The first technique is the Principal Component Analysis (PCA) while the second Is Linear Discriminant Analysis (LDA). In additi...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2013
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/11811/1/mt0000000641.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Face recognition is an important biometric application. In this thesis, a comparison of two common techniques for face recognition is carried out under the same conditions. The first technique is the Principal Component Analysis (PCA) while the second Is Linear Discriminant Analysis (LDA). In addition, the performance of PCA and PCA (with Radon) was also carried out. The Euclidean distance is used as the matching criteria. An investigation of the effect of the parameters of PCA on the performance of the face recognition system is carried out. First, it was found that the number of eigenvalues used affects the recognition rates of the system. The maximum number of eigenvalues used is 300. The equal correct rates increases from 1 until 40 to 80 eigenvalues used then become steady afterwards regardless of the image size used. Second, it was found that the higher the number of training images per person the lower the false acceptance rate. Third, the image size used effect the recognition rate when a fixed number of eigenvalues used. However, different image size has their own their optimum number of eigenvalues to achieve highest equal correct rate. When optimum eigenvalues used, their recognition rate did not vary significantly. A comparison of performance, time and resource used by all face recognition system is presented. Four individual systems are compared; PCA, PCA with Radon, LDA, and LDA with Radon. Each individual
system gives recognition rate of 89%,88%, 94%, and 92% respectively with LDA outperform the other three techniques. It was found that no improvement on recognition rate when PCA and LDA use the Radon Transform features as input
showing that applying Radon Transform on properly normalized frontal image does not boost the recognition performance. When compared the individual system to the data fusion system, it was found out that data fusion system gives better recognition rate than all the individual face recognition system. Fusion of PCA, LDA, and LDA with Radon give the best recognition performance, giving 98% correct recall and reject rate, and uses 62.8 second process time and 22.2 MB space. |
|---|