|
" Optimizing Computational Time of Face Recognition System Using Chinese Remainder Theorem "
Tajudeen, Madandola Niyi
Gbolagade, Kazeem Alagbe
| Document Type
|
:
|
Latin Dissertation
|
| Language of Document
|
:
|
English
|
| Record Number
|
:
|
1107770
|
| Doc. No
|
:
|
TLpq2455941256
|
| Main Entry
|
:
|
Gbolagade, Kazeem Alagbe
|
|
|
:
|
Tajudeen, Madandola Niyi
|
| Title & Author
|
:
|
Optimizing Computational Time of Face Recognition System Using Chinese Remainder Theorem\ Tajudeen, Madandola NiyiGbolagade, Kazeem Alagbe
|
| College
|
:
|
Kwara State University (Nigeria)
|
| Date
|
:
|
2020
|
| student score
|
:
|
2020
|
| Degree
|
:
|
Ph.D.
|
| Page No
|
:
|
189
|
| Abstract
|
:
|
One of the greatest challenges in the world is poor identification and verification of citizens and immigrants. It is of utmost significance to launch an identity of citizens to curb criminalities. There are several knowledge-based methods apart from human attributes methods referred to as biometrics. Biometrics is extremely trustworthy since they cannot be simply falsified, transformed or defrauded. Some of the universally enjoyed biometric types are the keystroke dynamics, ear, signature, face, fingerprints and voice. The choice of Face biometrics is as a result of its natural means of identification and does not require interference of suspect. Several studies on face recognition system shown that the computational time required by the Dimensionality Reduction (DR) algorithm is high. A good face recognition system apart from having high recognition accuracy must also be able to authenticate a person in lesser time. Facial DR algorithm consists of linear and nonlinear approaches. Principal Component Analysis (PCA) is an example of linear DR algorithm while Kernel Principal Component Analysis (KPCA) is one of the instances of nonlinear DR algorithm. PCA does not model difference between classes like other linear DR algorithms. Also, KPCA has merit of not specifying the number of components to be extracted in advance. Therefore, there is need to optimize the computational time and identification accuracy of KPCA and PCA dimensionality reduction algorithms by utilizing Chinese Remainder Theorem (CRT). PCA-CRT and KPCA-CRT algorithms were developed through employment of CRT with PCA and KPCA. Two databases were adopted for acquisition of datasets. Yale database was used as a standard database and a local database was setup. Yale database contains a total of 165 frontal pictures of fifteen persons with eleven face pictures of individual persons. In which one hundred and twenty face pictures were employed for training while forty-five pictures were employed for testing. Likewise, local database has a total of one hundred and twenty pictures of forty people frontal faces with 3 pictures of each individual. Eighty images was used for training while forty was used for testing. Recognition index, testing time and training time were performance metrics employed to determine the recognition correctness and computational time of KPCA and PCA with and without employing CRT with them separately. Euclidean distance was used as classifier. MATLAB 2015a Intel(R) Celeron (R) CPU with 1.60GHz Processor speed was used to implement the four algorithms (PCA, PCA-CRT, KPCA and KPCA-CRT). The experimental results were compared with the use of column and pie charts. T-test statistical tool was used to determine the statistical significance difference between KPCA and KPCA-CRT likewise between PCA and PCA-CRT. On local database, the average testing time of KPCA and PCA were 1.5475 seconds and 67.3016 seconds respectively. The experiment revealed total training time of 540.5134 seconds for PCA and 7240.5360 seconds for KPCA. Also PCA used 13.5128 seconds average training time while KPCA used 181.0134 seconds average training time. PCA has 72.5% performance recognition accuracy while KPCA has 80.0% performance recognition accuracy. The average testing time used by PCA-CRT was 1.5185 seconds while KPCA-CRT was 66.9082 seconds. PCA and PCA-CRT have equal performance recognition accuracy of 72.5% also KPCA and KPCA-CRT have equal performance recognition accuracy 80.0%. In the case of Yale databse, it was observed that PCA and KPCA has average testing time of 1.8093 seconds and 3.3276 seconds respectively. Total training time used by KPCA was 35206.4730 seconds while PCA used total training time of 1285.1600 seconds. The experiment revealed equal percentage of recognition accuracy performance for both PCA and KPCA both were able to recognise 32 images among the 45 images used as testing samples which was 71.1% despite the variations in some images recognized index in the database. PCA has 1.8093 seconds average testing time while PCA-CRT has 1.6863 seconds average testing time. Similarly PCA total training time was 1285.1600 seconds while PCA-CRT total training time was 1196.4553 seconds. Also, KPCA-CRT and KPCA have total training time of 33849.0530 seconds and 35206.4730 seconds respectively. PCA, PCA-CRT, KPCA and KPCA-CRT have recognition accuracy of 71.1%. The calculated T-test value of 5.9188 was greater than the T-test table value of 1.98 at 0.05 level of significance which shown that the difference between KPCA and KPCA-CRT computational time was statistically significant. The computational time difference between PCA and PCA-CRT was also statistically significant since the calculated T-test value of 7.0839 was greater than the T-test table value of 1.98 at 0.05 level of significance. The study revealed that testing time and training time used by PCA were far lesser than those of KPCA, that is KPCA has more computational time than PCA. Also, that employment of CRT reduced the computational time of the two algorithms with no effect on the algorithms recogniton accuracy.
|
| Subject
|
:
|
Artificial intelligence
|
|
|
:
|
Computer science
|
|
|
:
|
Statistics
|
| |