Study On Facial Image Synthesis And Digital Watermarking

Since a long time, facial image synthesis has been a hot research topic in the fields of computer graphics, image process and computer vision. Human has abundant and complex face expressions. It is a very challenging topic tosynthesize a human's facial image that is of special expression or age naturally and realistically. Natural and living facial images synthesized by computer have been widely applied to lots offields such as movie, amusement, virtual reality, videoconference, visual telephone, criminal detection, face recognition, computer interface and facial surgery, etc.With the development and dissemination of Internet, digital products are more and more abundant. Meanwhile, it also becomes more and more easy and convenient to copy digital information. This may lead to a big problem that digital products are abused illegally. So it is imminent to protest the copyright of digital products. Digital watermarking is a technique method for copyright protection, and hence turns into a hot topic in multiple media-safe fields.This paper reports our recent research works on facial image synthesis and digital watermark. Corel context includes image morphing, expression synthesis, aging simulation and digital watermarking. Novel ideas and improvements are proposed by analyzing the technique details in the following four issues.1. Image Morphing. Image morphing is a basic method as a tool of facial image synthesis. Aiming at abnormal phenomena of image morphing produced by the method based on Gaussian radial basis functions in common use, two different improvements are proposed for settling the problem effectively.The conception of "asymmetric radial basis function" is proposed firstly, and the asymmetric form of Gaussian radial basis function is presented. Then a novel boundary stabilized image-morphing algorithm based on asymmetric radial basis functions is proposed. Asymmetrical radial basis functions can control asymmetrical range and hence can keep the boundary of a morphed image stable if the image-morphing algorithm uses such asymmetrical bases. It is a highlighting character differing from symmetrical radial basis functions that are used for traditional image morphing algorithms in which over warping on image boundary often occurs. This asymmetricalmethod is very effective;the morphed images obtained by the asymmetric method can give good local morphing results and have lubricities as well as that produced by algorithms based on symmetrical radial bases. More importantly, it can stabilize the image boundary.A locally constrained offset based image-morphing algorithm is proposed. This method is effective for local constrained image morphing and it can automatically reduce unnecessary offset outside the character part produced by the method using symmetrical Gaussian radial basis functions. This method is also simple;it does not need manned work on setting control radius.2. Facial Expressions Synthesis. We explore how to synthesize a natural and lively facial expression image from a facial neutral image and multiple men's facial neutral images and facial expression images.A synthesizing method for image morphing based facial expression is proposed. This method can be easily realized since its base ideas are simple and its produces are convenient, too. At one time, we propose relative offsetting that can match different face figures automatically. We also propose the concept and method of "expression details template" that can enhance facial expression. These build the bases of synthesizing more expressive and lively facial expression images. Facial expression images produced by this method are quite natural and lively. Most importantly, this method can produce different facial expressions by adjusting weight vectors used in this method. We can obtain more general and abundant facial expression image if there are enough deferent-mode expression images.We propose a morphable model that bases facial expressions synthesizing. This method gains shape and texture information of the facial expression image synthesized based on linear combination approach, the basic idea of facial image morphable model. This method can be easily realized within simple produces. Facial expression images produced by this method are quite natural and lively. Most importantly, this method can produce facial expressions with mouth opened and teeth showed.3. Face Aging Simulation. We explore how to simulate a natural and lively aged face of a youngster, using multiple face-images of elders as consulting targets.Two novel methods for face aging simulation are proposed. Based upon the morphable model, the linear combination properties of the youth-face to the possible youth-images of the elders are first extracted and then kept in constructing an agedface-image as required. Because of its linear property, Gaussian low-pass filtering is used to construct the possible youth-images of the elders. An iteration approach is presented to improve the simulation results for aging simulation. This iterative method also works for single elder's image. We also propose a flexible method by weighting the multiple simulations that are obtained respectively by the iterative method with single consulting target. The methods can produce more natural simulation results for aging a youth-face image, and because of multiple older-face images used as synthesizing sources;the resulting image absorbs and fuses multiple aging modes. There are no constraints on poses and lightings for the consulting images or the youth-face image.4. Digital Watermarking. We develop novel digital watermarking technique based on new ideas.A SVD based two-layer watermarking algorithm is proposed. It can embed watermark into two different layer. This algorithm has a good transparence of the embedded watermark image and is robust against to attacking. It's important that it can extract out the watermark embedded in another layer when one layer watermark is removed or replaced by another watermark.A matrix-norm based watermarking algorithm is proposed. It's presented after analyzing the essence of the SVD based watermarking algorithms. Basically it is based on the selected matrix-norms of sub-images and their quantization. This algorithm is quite simple and has relatively low cost compared with SVD-based algorithms. This algorithm has a good transparence of the embedded watermark image and is robust to attacking.