Performance Evaluation of an Improved Self-organizing Feature Map and Modified Counter Propagation Network in Face Recognition

Abstract

Aim: To carry out performance evaluation of an Improved Self-Organizing Feature Map (SOFM) and Modified Counter Propagation Network (CPN) techniques in face recognition. These two techniques were examined, implemented and evaluated by using metrics such as recognition accuracy, sensitivity and computation time. Problem/Study Design: In lieu of threat to global peace and criminal activities in our society today, it is then imperative to adopt a non-linear techniques that might improve the recognition performance of face recognition system because of their intrinsic characteristics. A comprehensive evaluation of these two selected artificial neural network techniques was performed to address these challenges and to estimate the preferred technique that had manifested an improved system. Place and Duration of Study: Ladoke Akintola University of Technology (LAUTECH), Ogbomoso, Nigeria and was done during the period of the Master Study. Methodology: An Africa database of 240 face images was created by capturing six face images from 40 persons with a digital camera. Image pre-processing was carried out using MATLAB and normalized using local histogram equalization for contrast enhancement. Principal Component Analysis (PCA) was used to extract distinctive features and reduce the dimensionality of each image from 600 x 800 pixels to four different dimensions; 50 x 50, 100 x 100, 150 x 150 and 200 x 200 pixels. SOFM and CPN techniques were used as classifiers for face recognition then evaluated using 140 images for training and 100 images for testing with best selected similarity threshold value. The two techniques were evaluated using recognition accuracy and computation time as performance metrics. Results: The results of evaluation showed that, at 50 x 50 pixels, SOFM had 81% accuracy with computation time of 243 s while CPN gave 84% accuracy in a time of 174 s. Correspondingly, at 100 x 100 pixels, SOFM had 83% accuracy with a time of 244s whereas CPN had 88% accuracy with a time of 179 s. Similarly, at 150 x 150 pixels, SOFM gave accuracy of 87% with a time of 245 s while CPN generated 90% accuracy with a time of 190s. Furthermore, at 200 x 200 pixels, SOFM resulted in accuracy of 92% with a time of 249 s, however, CPN had 95% accuracy with computation time of 234 s respectively. Conclusion: This research has shown that CPN outperformed SOFM techniques in face recognition based on recognition accuracy and computational time.