DEVELOPMENT OF AN EFFICIENT FACE RECOGNITION SYSTEM BASED ON LINEAR AND NONLINEAR ALGORITHMS

Abstract

Face biometric is becoming primary biometric technology because of rapid advancement in the technology such as digital cameras and mobile devices which in turn facilitates its acquisition and usage. The problems of face recognition, which include but not limited to, arbitrary pose, lighting and facial expression, are very difficult and unsolved problems. Although some algorithms have been developed, the algorithms are not at its optimum. To cope with these problems, face recognition systems have to utilize recognition techniques capable of extracting stable and discriminative features from facial images regardless of the conditions governing the acquisition procedure. This research work is motivated by the need to establish the most efficient algorithm among several that has been developed based on standard parameters. The first stage involves the preprocessing of images by the use of Gabor Wavelet for image representation to exhibit desirable characteristics of spatial locality and orientation selectivity that are optimally localized in space and frequency domains. Linear and nonlinear face recognition dimensional reduction techniques are then used to project the feature space into lower-dimensional space prior to the matching stage. In the final stage, Mahalanobis Cosine metric is used to define the similarity measure between two images. It is the cosine of the angle between the images after they have passed through the corresponding dimensional reduction techniques. Lastly, performance analysis of the linear and nonlinear algorithm were carried out (with and without Gabor for each algorithm) using three performance metrics: Cumulative Match Score (CMC), Receiver Operating Characteristic Curve (ROC), and Expected Performance Curves (EPC). The America Telephone and Telegraph Corporation (AT & T) Laboratories Cambridge data base is used for the work. It contains 400 images of 40 different subjects. 120 images are used for training. 160 images are selected for testing and another test set containing 120 images called the evaluation set is provided for more evaluation and for plotting the Expected Performance Curve that required two data sets. The experiments show that Linear Dicriminant Analysis (LDA) and Kernel Fisher Analysis (KFA) have the highest performance of 93.33% by incorporating Gabor wavelets. LDA outperformed other methods in most cases but have a very close recognition rates with KFA. Kernel Principal Component Analysis (KPCA) performed worst with a recognition rate of 49.29% without the use of Gabor wavelets and 56.88% with the use of Gabor wavelets. The result revealed that the performance of the linear and nonlinear algorithms depend on the nature of tasks at hand such as the number of classes of the images, preprocess of images, and the number of probe images. The result of the work shows that the linear algorithms perform better and are more capable of excellent performance in a variety of realistic biometric scenarios.