Adaptive Confidence Level Assignment to Segmented Human Face Regions for Improved Face Recognition

Abstract

Improving the existing face recognition technology to a higher level and to make it useful for many areas of applications including homeland security is a major challenge. Face images are prone to variations that are caused due to expressions, partial occlusions and lighting. These facial variations are responsible for the low accuracy rates of the existing face recognition techniques especially the ones that are based on linear subspace methods. A methodology to improve the accuracies of the face recognition techniques in the presence of facial variations is presented in this paper. An optical-flow method based on 'Lucas and Kanade' technique has been implemented to obtain the flow-field between the neutral face template and the test image to identify the variations. Face recognition is performed on the modularized face images rather than the whole image. A confidence level is associated with each module of the test image based on the measured amount of variation in that module. It is observed that the amount of variations within a module is proportional to the sum of the magnitudes of the optical-flow vectors within those modules. Least confidence is attached to those modules, which has the maximum sum of magnitudes of the optical-flow vectors. A K-nearest neighbor distance measure is implemented to classify each module of the test image individually after projecting it into the corresponding subspace. The confidence associated with each module is taken into consideration to calculate the total score for each training class for the classification of the test image. Analysis of the algorithm is performed with respect to two linear subspaces - PCA and LDA. A high percentage of increase in accuracy is recorded with the implementation of the proposed algorithm on available face databases when compared with other conventional methods