3D Face Reconstruction Based On A Single Frontal Face Image

Technology of 3D face modeling has become many researchers'hotspot at home and abroad along with the rapid development of computer graghics and 3D digital technology and a lot of achievements have been made so far. However, there are still some problems and faults in this field. First most of 3D face reconstruction systems still need users'manually selecting of face feature point at present. Then on the one hand the feature points may vary from different persons and on the other hand system's automation degree will lower down because of this. Secondly most of the system's efficiency is still low and its expansion to other fields is blocked in a certain extent. Therefore this paper proposed a simple and rapid method of 3D face modeling with high realism based on a single frontal face image in the condition of neutral expression and uniform illumination through automatic extraction of the face feature points.This paper first compared the advantages and disadvantages of current typical 3D face modeling systems. According to the characteristics of the topic and the existing experimental condition, CANDIDE-3 face model was selected. The 26 key feature points were selected for 3D face modeling with the vertices of CANDIDE-3 model as reference and combing the MPEG-4's definition to face feature points.Face was detected by using the AdaBoost method to remove complex backgraound caused by filming under natural conditions and thus the search range can be narrowed for the accurate extraction of face feature points. Then in the determining face region rough range of eyes, lips and nose can be determined by using gray integral projection method based on prior knowledge of facial feature distribution. According to the circular geometric characteristic of pupil, this paper proposed a fast and effective approach to locate the pupil accurately based on Hough transform to detect circle and thus the other feature points on the eyes'contour can be selected. According to the differences between skin color and lip color, this paper proposed an algorithm with stronger anti-interference ability to separate the lip based on the Red Exclusion algorithm joining the distribution difference between R component and G component and thus the feature points of lip can be selected. Nose can be separated considering nostril's gray characteristic and thus the feature points of nose can be selected. The experiments prove that both detection accuray and detection speed are satisfying.A target geometric model approximate to the face in the image was constructed by modifying CANDIDE-3 model based on the key feature points detected before. Then an individual realistic 3D face model was generated through texture mapping to the target geometric model.On the basis of above-mentioned theory reaserch, a realistic 3D face automatic modeling prototype system was designed and realized with VC++6.0 by using OpenCV and OpenGL. The experiments show that the method of this paper is simple and fast. The system has high efficiency and the result is realistic. Face feature points are auto-extracted by computer and thus automation degree is high.