Novel elastic graph matching-based face authentication with optimal class separability

In order to resolve the problem in traditional elastic graph matching that the cost function cannot effectively reflect the facial shape difference, and the weights for cost components cannot be determined automatically, a novel face authentication algorithm with support vector machine-based elastic graph matching is presented. This algorithm first introduces facial geometric feature similarity cost in the cost function, which enhances its discriminant power on shape difference. Then it takes all cost components as a similarity vector and trains the support vector machine with training samples. Finally it generates the new cost function through the classifier's discriminant. The new cost function reflects the relative importance of each cost component, and has optimal class separability. Experiments show that this algorithm gets 6.6% improvement in face authentication rate, and 1.5 pixel of decrease in dynamic matching error.