Research On Human Face Synthesis And Visual Fatigue Detection Based On Facial Landmark Locating

With the rapid development of computer vision technology, human facial information processing plays an important role in fields of human face synthesis and visual fatigue detection. The factors like illumination, pose and expression variations make human face image more complicated. Accurate locating of human facial landmarks can improve the effect of multi-pose human face synthesis and visual fatigue detection. Therefore, we focus on multi-pose human face synthesis and visual fatigue detection based on human facial landmark locating in this thesis. The main achivements are summerized as follows.To address the problems of inaccurate locating of facial landmarks in the mixtures of tree model and missing detection of faces in Supervised Descent Method(SDM), we propose a multi-pose human face detection and facial landmark locating algorithm by combining the mixtures of tree model with SDM. We initialize the localization of multi-pose faces by the mixtures of tree model, then locate and refine the landmarks of inner facial contour by SDM, which are combined with the outer facial contour to achieve all facial landmarks of multi-pose faces accurately.To solve nonlinear representation of the pose subspace in multi-pose human face synthesis, we propose a modified multi-pose face synthesis algorithm. We apply tensor decomposition on the shapes of multi-pose faces in the training set to separate the pose and identity subspaces. We get the pose manifold by spline fitting in the pose subspace. The identity coefficient of the test image is synthesized through sparse representation in identity subspace. Based on the pose manifold and the synthesized identity coefficient, we obtain the shape manifold of the new identity. Aimed at image stretching in the corner of eyes and nose caused by pose variation, we add four additional landmarks using the geometrical relationships of facial landmarks to modify the results. Finally, we warp the texture of the test frontal face image to the shape manifold through affine transformation to synthesize the multi-pose faces of the test image. The synthesized multi-pose face images in our method maintain the shape and identity of the test image very well. The synthesized facial texture looks natural and real.Since fast and accurate facial landmark locating is urgent in visual fatigue detection based on human face, we initialize the localization of human face by Adaboost based human face detector, then localize and track human facial landmarks in the video by SDM. We set a series of rules for visual fatigue detection according to the ratio of eye area to eye distance and the aspect ratio of mouth based on the facial landmarks. The experimental results on driving videos and simulation videos show the superiority of our method in visual fatigue detection.