Research On3D Face Reconstruction Based On Morphable Model

The3D face reconstruction is one of the core research fields of computer vision,it involves the application of3D face recognition,3D face animation,3D games,simulation, virtual reality etc. In recent years, with a large number of facereconstruction algorithm emergencies,3D face reconstruction has achieved goodresults in many application fields, such as medical, criminal investigation, games,movies, animation and other fields. As the practical application of modelreconstruction requires continuous improvement of the accuracy, efficiency of thealgorithm, the degree of automation, to improve the performance of the algorithm inthese three aspects has been the research target.Compared with the muscle model, parameter model, visual model, etc., themorphable model is currently the best3D face reconstruction algorithm, which hasthe characteristics of good sense of reality, high degree of automation. The morphablemodel has the advantages of the following two points when compares to the purereconstruction algorithm based on images. First, the reference face usingthree-dimensional morphable model to establish a probability distribution model,which overcome the problem of non-face reconstruction "Shape from X" methodappearing. Second, morphable model transfers the three-dimensional reconstructionproblem into a series of optimization parameters by using a priori knowledge, whichcan tolerate light changes, even in the case of facial sweating can realistic modeling.Although this model has unparalleled advantages than other algorithms morphablemodel, it also has some problems. The first one is high computational complexity andslow convergence; the other one is modeling algorithm mainly focusing on theadjustment of the model parameters and the low degree of automation. Therefore,people rely on the reconstruction process either manually adjust or the initial need tomatch the picture above when tagging a number of characteristic points.Aiming at the existing problems of morphable model, the work of this paperfocuses on two aspects. Firstly, this paper analyzes the strengths and weaknesses oforiginal morphable models by using stochastic gradient descent algorithm in thematching stage. And because the algorithm converges slower flat and easy to fall intolocal minima, this paper attempts to introduce a method by designing conjugategradient algorithm based on random direction to accelerate the iterative convergencerate. Secondly, for the degree of automation of the existing algorithms is not high, thisarticle will introduce facial feature detection algorithm giving external parameterdeformation process of matching model initialization after the features location offace, nose and eyes. Through this process, the original manual work will be done bymachines. In extreme cases, when the positioning failure of the face deflection in45degrees, eyes and nose, this algorithm can still use manual calibration. This flexiblechoice not only makes sure the algorithm accuracy and execution speed, but alsoimproves the automation degree.