AAM And Tensor Decomposition Based Multi-view Facial Image Synthesis Algorithms

With the rapid development of Computer Science, the technique of face synthesisis playing a more and more important role in the fields of Artificial Intelligence andPattern Recognition. So the synthesis of a lifelike face image is a hot research topic.However the non-rigid property of the face structure causes the influence on faceimages by multi-factors. Among them, the variation of views causes a serious problem.With the further development of the face synthesis techniques, synthesis of themulti-view face images has become an important and challenging issue. Fortunately,statistical learning based face synthesis has becoming a popular direction.In this paper, the algorithm of synthesizing multi-view face images has beendiscussed deeply. Firstly, to handle the error caused by the undersampling problem ofthe high dimension data, the approach combining Active Appearance Model (AAM) andTensor analysis theory is proposed for separating facial influence factors. With thefeature points abstracted by AAM and factors separation produced by Tensor algorithm,we solved the undersampling problem on the precise description of the multipleinfluential factors on face images.Secondly, the approach to synthesize the facial feature points under continuousviews is proposed with the combination of Sparse Representation and ManifoldModeling. Through the reconstruction of the identity information of the training imagesand the fitting of the new view information, this approach could synthesize the locationof feature points of the test image under different views. Experimental results show thatthe proposed approach could synthesize the feature points under the views absent intraining set with the test data in different views.Finally, the synthesis method of frontal face image based on the global-localtransformation with some occluded facial information is proposed. The double imagetransform strategy based on different information of the facial regions is applied tosynthesize the frontal face image corresponding to the input image. The experimentalresults show that the proposed approach could synthesize the vivid frontal face imageseven when the input image has large rotation angle and obviously occluded texture.