Study Of Facial Feature Point Detection For Multi-pose Face Based On Cascade Regression

Face image contains lots of information, like facial expression, head pose, identity. To extract the information, facial feature point detection is an essential step. Facial feature point detection is an automatic facial calibration algorithm, which aims at extracting a group of points that can be used to describe the facial structure of face image. However, in complex environments, affected by pose variance, the performance of traditional facial feature point detection algorithms would decline significantly. To solve this problem, an improved facial feature point detection algorithm based on cascade regression is proposed in this paper.The major processes of facial feature point detection algorithm based on cascade regression are as follows. First, the initial position of facial feature points is set. And then, the pose-indexed features are extracted from the face image with the position of feature points. These features are used as input of a weak regressor. Then the result of weak regressor updates the position of feature points. These processes are repeated thousands of times to locate the precise position of facial feature points.To improve the robustness of cascade regression algorithm while pose changing, a new pose-indexed feature, Average Pixel of Local Area feature, is proposed in this paper. The feature divided the face image into many local areas according to the position of feature points. Several reference points are selected in every local area. The average pixel value of reference point are set as the output of each local area. And the Average Pixel of Local Area feature is defined as the diff of local areas output. Experimental result shows that the detection precision improves by using the feature proposed in this paper.To solve the problem that the facial feature point detection based on cascade regression is sensitive to the initial position of facial feature points, a facial feature point initialization algorithm based on dynamic programming technique is proposed in this paper. The algorithm extracts feature from every feature point, and selects the fittest initialization result in a large number of candidate results with the help of dynamic programming technique. The experimental result shows that the detection algorithm is less affected by the variance of pose with proper initial position that found by the algorithm proposed in this paper, which lead to an improvement of feature point detection precision.