Normalized Independent Component Analysis for Face Recognition

Aims: To design a Face Recognition System (FRS) using combination Independent Component Analysis (ICA) and Artificial Neural Network - Normalized ICA. In order to improve the performance of a conventional ICA which suffers the drawback of ranking the energies of the generated features.

Study Design: The FRS was simulated using Matlab 2011 version. An algorithm was developed which combines the ability of the conventional ICA with ANN to generate final predictions. The ANN serves as a region finder and generated likely predictions associated with face image classes. Hence, reduced search space during testing.

Place and Duration of Study: Department of Computer Science and Engineering, Ladoke Akintola University of Technology, Ogbomoso, Nigeria, between June 2014 and July 2015.

Methodology: 40 individuals face images were captured in an uncontrolled environment. The face database comprises of 10 images for each subject taken at different times. The images were pre processed by cropping to different sizes (92*92; 92*100; 92*112 pixels respectively) and removing unwanted background. During testing Euclidean distance was used as similarity measure and faces were classified as “known” if less or equal to the threshold value set else “unknown”.

Results: The recognition accuracies at dimension 92*92 are 86.00% and 95.00% for ICA and NICA-based system using 30 principal components, 86.50% and 96.00% using 60 principal components at the same dimension respectively. At dimension 92*112 a recognition accuracy of 90.00% and 98.00% was obtain for ICA and NICA-based system, 91.00% and 98.00% using 60 principal components at the same dimension respectively. At cropped dimension 92*92 it took an average of 0.0096s and 0.0095s using 30 principal components to recognize a test image in ICA-and NICA-based, 0.0086s and 0.0085s using 60 principal components at the same dimension respectively and at cropped dimension of 92*112 it took an average of 0.0102s and 0.0098s using 30 principal components, 0.0106s and 0.0099s using 60 principal components at the same dimension respectively

Conclusion: The developed NICA-based system has better recognition accuracy than a conventional ICA-based system and also recognizes face images faster.