Facial Expression Action Unit Recognition System

The recognition of facial expression action unit is the most important component in current expression recognition field. The recognition can be divided to two parts: the first part is face detection, and the second part is facial expression recognition. Aim at the problem of low performance in face detection and the limitation in multi-pose detection, we used a new feature algorithm named NPD based the cascading Ada Boost classifier. The result show that this algorithm detecting human face faster than Viola-Jones algorithm, also handle multi-pose face detection. Because of the enormous number of global features, facial expression was slow. We proposed a block-base facial expression action unit recognition algorithm. With the algorithm, we can divide Action Unit to two classes, the first is eye-position-based, and the other is mouth-position-based. The result of experiment shows that the speed is faster than global features and more accurate. Aim at the high algorithm complexity of Gabor filter, we proposed a CUDA-based Gabor filter algorithm. Compared to CPU-based Gabor filter, can increase the speed by almost two order of magnitude.