3D facial point cloud preprocessing based on skin color detection using SVM

3D face recognition has gained increasing research attention among researchers in recent years. In comparison to its 2D counterparts, 3D face recognition systems have the potential for better recognition accuracy and robustness. In 3D face recognition, it is necessary to extract pure facial part in the point cloud to improve accuracy, which was mainly conducted manually in most previous studies. This paper proposed a fully automatic approach for 3D face point cloud preprocessing, which can extract faces in uncontrolled environments, such as large pose variations, partial occlusions and changes of expressions. Considering that 3D structure of the face can often be sensed simultaneously with color information, this paper uses color information to extract pure face part of the scan. An SVM model with skin and non-skin colors is trained and skin points in 3D facial point clouds are detected using the trained model. To fill holes on the obtained skin point cloud, This paper proposed an approach based on KNN and used a threshold to determine whether a point is on the face surface or not, such we can fulfill the holes on faces and removing outliers. Experiments show that the proposed algorithm can extract faces of different scales, poses and expressions under various illumination conditions stably and accurately.