Face Image Synthesis Based On Key Points Location

Augmented Reality(AR)technology refers to superimposing a computer-generated virtual model into the real physical world that the user can perceive.It is a new type of development based on Virtual Reality(VR)Computer application technology.Augmented Reality technology better combines the physical world with the virtual world.So far,many AR applications have been applied to military,medical,education,entertainment,architecture and other industries.The face synthesis technology proposed in this paper provides a new research direction for augmented reality.This face synthesis technology has not only been able to complete a simple face-to-face or face style conversion,but is able to synthesize a new face that does not exactly match the real people.The result of the algorithm synthesis will fuse two input face features,which is more inclined to the seamless integration of human face details,and solves the problem of background ghosting commonly found in previous methods.It can be expected that this combination of face synthesis technology and augmented reality will certainly provide new ideas for the development of AR applications.The face synthesis algorithm proposed in this paper mainly includes the following steps:face detection and face landmarks detection,face fusion,face swapping.The main work and research results of this paper are as follows:(1)In the phase of face landmark detection,In the key-point location phase of the face,the advantages of the GBDT[61]algorithm in key point location are analyzed in detail,and successive iterative approximation are used to detect face landmarks.This method uses different initialization values to obtain different regression objectives to expand the data set.The cascaded regression is used in this method.Each stage of the re-initializer re-allocates the initial value of each sample,which greatly improves the accuracy of the algorithm.Finally,we conducted comparative experimental analysis of different regression series and different sample initialization methods and other algorithms,which proved that the key point location method used in this paper has higher accuracy.(2)In the face fusion stage,meshes are generated on the two input images,and the face landmarks of the fusing image are obtained based on the positions of the face landmarks of the two input images.Then,we analyze the advantages and disadvantages of several interpolation methods,and apply affine transformation and bilinear interpolation methods to image deformation.(3)In the phase of face swapping,a replacement-based face swapping method was proposed.This method can make the final result consistent with the original face's face shape through the deformation of the reference image,which is of great significance to the application of the face change algorithm.A face semantic segmentation algorithm is introduced to accurately segment the ROI region of the human face and solve the problem of inaccurate reality caused by occlusion of the human face in the image.(4)At the stage of image synthesis,a new image synthesis process is proposed.The method in this paper ensures that the synthetic image has a high degree of realism,and the synthetic result has more ornamental value and application value than the face synthesis in the usual sense.