Face Detection Algorithm Research Based On Color-Model

Face recognition is one type of biometric technology. A fundamental and key issue in face recognition is face detection. The task of face detection is detecting out human face from the image background. The different image backgrounds, diverse variation of face appearance and the varying illumination condition lead to difficulties for face detection. Therefore, face detection is a very challenging problem. Based on static human face detection in color image, the skin color features and relative detection algorithms are investigated in this paper.An enhanced face detection algorithm based on the skin color and multiple template matching is presented and implemented in this paper. Firstly, skin color segmentation is performed on the color image after pre-processing by the two-threshold. Then, matching search is performed based on multi-template matching. Finally, mosaic rule is used to verify. Experiment results show that the influence of the variation of face appearance on the detection performance of the proposed algorithm is very small. And the proposed method has achieved high detection speed and is easy to implement on testing a large number of images.In order to solve the problem of the great computational complexity and the low speed of the Bayesian Discriminating Features (BDF) algorithm, a algorithm combining the skin color with BDF algorithm is presented and implemented in this paper. Firstly, the skin color segmentation is performed to obtain one or more face candidates. Secondly, Discriminating features and statistical model are obtained for the trained average face and average non-face model. Then the BDF method is used for the determination of the face candidates. Furthermore, in order to improve the detection performance, the non-face regions with very high error detection ratio is carried on trains again. Experiment results show that: in contrast to the original BDF algorithm, the detection speed of the enhanced BDF algorithm is improved 1.6 and error detection ratio reduced by 1%.