FACE CLASSIFICATION FOR AUTHENTICATION APPROACH BY USING WAVELET TRANSFORM AND STATISTICAL FEATURES SELECTION
This thesis consists of three parts: face localization, features selection and classification process. Three methods were proposed to locate the face region in the input image. Two of them based on pattern (template) Matching Approach, and the other based on clustering approach. Five datasets of fac...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2011
|
| Online Access: | http://utpedia.utp.edu.my/3053/1/Nadir_Norain-final-FINAL_thesis.pdf http://utpedia.utp.edu.my/3053/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This thesis consists of three parts: face localization, features selection and classification process. Three methods were proposed to locate the face region in the input image. Two of them based on pattern (template) Matching Approach, and the other based on clustering approach. Five datasets of faces namely: YALE database, MIT-CBCL database, Indian database, BioID database and Caltech database were used to evaluate the proposed methods. For the first method, the template image is prepared previously by using a set of faces. Later, the input image is enhanced by applying n-means kernel to decrease the image noise. Then Normalized Correlation (NC) is used to measure the correlation coefficients between the template image and the input image regions. For the second method, instead of using n-means kernel, an optimized metrics are used to measure the difference between the template image and the input image regions. In the last method, the Modified K-Means Algorithm was used to remove the non-face regions in the input image. The above-mentioned three methods showed accuracy of localization between 98% and 100% comparing with the existed methods. In the second part of the thesis, Discrete Wavelet Transform (DWT) utilized to transform the input image into number of wavelet coefficients. Then, the coefficients of weak statistical energy less than certain threshold were removed, and resulted in decreasing the primary wavelet coefficients number up to 98% out of the total coefficients. Later, only 40% statistical features were extracted from the hight energy features by using the variance modified metric. During the experimental (ORL) Dataset was used to test the proposed statistical method. Finally, Cluster-K-Nearest Neighbor (C-K-NN) was proposed to classify the input face based on the training faces images. The results showed a significant improvement of 99.39% in the ORL dataset and 100% in the Face94 dataset classification accuracy. Moreover, a new metrics were introduced to quantify the exactness of classification and some errors of the classification can be corrected. All the above experiments were implemented in MATLAB environment. |
|---|