A Master Thesis Submitted To University Of Electronic Science And Technology Of China

Face replacement in image means replacing the face region of a target face image with the face region of a source face image. The technology of face replacement has great value in entertainment, virtual reality, privacy protection, video chat. In this paper, we propose an improved method for automatic face replacement in images and realize an automatic face replacement system. The processing speed and the result of the face replacement system are desirable. This paper mainly includes the following research work:Firstly, we propose a simple and effective method of head pose estimation in face replacement for reducing the complexity of head pose estimation. Human head posture can be described by pitch, yaw and roll. Since pitch and yaw is difficult to accurately calculate, this article does not calculate the two angles directly. We just calculate roll. According to the "three court five eye" theory and the positions of facial feature points, we use pitch distance scale, yaw distance scale and roll to estimate the head pose.Secondly, we implement a method for face replacement under glasses occlusion in images. Before blending a source face image into a target face, we should detect glasses. When replacing the face image, the obstructions in face region are reserved, only replace mouth, nose, eyes and eyebrows. If there is no obstruction, we replace the face region directly. First, we detect glasses in face region by training a glasses detector classifier. According to the result of glasses detection, we select an appropriate face replacement method to replace the face of a target face so that the result of face replacement looks more nature and realistic.Thirdly, we implement a method for real-time video face replacement to solve the deficiency in the speed of traditional face replacement. First, detect face in the frames of video. After a face is detected, the face region is set as a region of interest. Then, extract 68 facial feature points: 10 points on two eyebrows, 12 points on eyes, 11 points on nose, 18 points on mouth, 17 points on face contour. Then, use these face feature points to align the source face image and the target face image. Finally, use poisson image fusion algorithm to blend the source face image into the target face image seamlessly.We implement a real-time automatic face replacement system. The experiments show that the system can achieve good result when the target face has glasses occlusion. The speed of the system is real-time. Compared with the traditional automatic face replacement techniques, it is faster and it does not require 3D facial reconstruction which is complex and time-consuming. It has great theoretical and practical value.