Automatic Facial Feature Point Localization Based On Similarity

Technology of automatic facial feature point localization is a hot spot of research in the field of 3D face modeling, facial expression recognition, etc. Currently, the number of facial feature point orientation method is large, but the improvement still needed on the calculation speed and detection effect of the algorithm.This paper introduces a kind of method based on the similarity between the skin color model and test images in the Cb-Cr subspace. This method select irises center, the corner of eyes, the eyelid points, the nostril points, the tip of the nose and mouth center, corner of lips and the edge of eyebrows points and the profile points of face, a total of 29 points were automatic located. The characteristics of this method are following: In a complicated illumination, beard interference and a small angle tilt of face condition, this system which mentioned in this paper is still robust. It is not necessary to localize the facial feature points manually to train them as a sample set.First of all, the original image is preprocessed. It includes histogram equalization and light compensation. The experiment result shows that the color image after preprocessing is clearer and brighter than the original image, and the contrast is enhanced. It makes that it is convenient for the following steps of face detection and facial feature point localization.Second, the preprocessed images are converted from the RGB color space to the YCbCr color space. Then, the similarity between the skin color model and the preprocessed images are calculated in the Cb-Cr subspace. The similarity calculated gray-scale images make the face area brighter and the non-face region darker. After using average filter to reduce the noise points of it and normalize it, the binary images which segmented the face area can be calculated. In the binary images, the morphologic calculation is adopted for removing noise.Third, Applying the binary images mentioned above and combining the prior-knowledge with edge detection algorithm and lip color formula, nostril gray-scale transformation function and eyebrows gray-scale transformation function to locate eyes, nose, lips and eyebrows feature points. One of the innovations in this project is the utilization of nostril gray-scale transformation function and eyebrows gray-scale transformation function.The method of locating the six face profile points is also the innovation in this paper. Based on the y-coordinates of the corner of eyes and the corner of lips and the x-coordinate of the center of two eyes, the five points except chin point can be easily located in the edge detected images. And the chin point can be detected by applying the curve fitting algorithm and adjust the y-coordinate of it by an angle which mentioned in the following text.At last, the experimental results verify the feasibility of this method used in this project. It is performed accurately in face detection and facial feature point localization.