HGOAAM: Facial Analysis by Active Appearance Model Optimized by Hybrid Genetic Algorithm
Résumé
Facial analysis is one of the challenges of a Human Machine Interactive (HMI) system. It includes face alignment, pose estimation, face recognition, facial expression detection etc. This paper proposes an efficient optimization technique by the hybridization of genetic algorithm (GA) with gradient descent (GD) to make a robust, efficient and real time face alignment system. It employs 2.5D Active Appearance Model (AAM) for the face search, in which the face model is generated by taking 3D landmarks and 2D texture of the facial image. Facial large lateral movements requires to optimize 6DOF (Degrees of Freedom) pose parameters which make the facial search space of AAM non-convex. This non-convex multidimensional search space requires an efficient optimization methodology. Gradient descent also known as deterministic optimization method can optimize the parameters as long as the search space remains convex. However in non-convex search space GA is able search face globally while GD can help GA to search face locally. In other words exploitation property of GD and exploration property of GA are combined to make an efficient optimization method. For this hybrid optimization we propose a gradient operator in GA, which functions in conjunction with the existing genetic operator of mutation. Thus it does not increase the computational cost of the system. We compare it with classical search algorithm by gradient descent and the hybrid optimization of GA with Simplex. Facial databases of SUPELEC'08, Pointing'04 and synthetic characters comprising of different facial poses are analyzed. Simulation results validate the efficiency, accuracy and robustness achieved.