Performance Evaluation of Feature Extraction Techniques in Multi-Layer Based Fingerprint Ethnicity Recognition System

This paper is set out to evaluate the performance of feature extraction techniques that can determine ethnicity of an individual using fingerprint biometric technique and deep learning approach. Hence, fingerprint images of one thousand and fifty-four (1054) persons of three different ethnic groups (Yoruba, Igbo and Middle-Belt) in Nigeria were captured. Kernel Principal Component Analysis (K-PCA) and Kernel Linear Discriminant Analysis (KLDA) were used independently for feature extraction while Convolutional Neural Network (CNN) was used for supervised learning of the features and classification. The results showed that out of sixty (60) individual fingerprints tested, eight (8) were classified as Yoruba, forty-eight (48) as Igbo and four (4) as Hausa. The Recognition Accuracy for K-PCA was 93.97% and KLDA was 97.26%. For Average Recognition time, K-PCA used 9.98seconds while KLDA used 10.02seconds. The memory space utilized by K-PCA was 94.57KB while KLDA utilized 52.17KB. T-Test paired sample statistics was carried out on the result obtained; the outcome presented reveal that KLDA outperformed the K-PCA technique in terms of Recognition Accuracy. The relationship between the average recognition time () and threshold value () was found to be polynomial of order four (4) with a high correlation coefficient for KPCA and polynomial of order three (3) with a high correlation coefficient for KLDA. In terms of computation time analysis, KLDA is computationally more expensive than KPCA by reason of processing speed.