【作者】 张浩；

【导师】 张举勇；

【作者基本信息】 中国科学技术大学 ， 计算数学， 2021， 硕士

【摘要】 随着计算机技术的迅速发展,视频通话和视频会议越来越普遍。在一些通讯设备中,由于摄像头角度问题,视频中的人脸呈现出俯视或者仰视的角度,将这种有角度的人脸矫正为平视的人脸成了一种新的需求。人脸矫正的一种思路是从视频中每一帧的图像中重建出人脸的三维模型,将这个俯视或者仰视的三维人脸模型旋转成正视的模型,再将其旋转再投影到二维平面上。在这种思路的基础上,有两个问题需要解决,一个是从单张二维人脸图像中重建出精度高、稳定性好的三维人脸模型,满足人们的视觉需求。另一个是遮挡带来的问题。人脸被遮挡区域的关键点检测会不准确,影响三维人脸重建的效果,进而影响人脸矫正的效果。所以在矫正的过程中我们需要判断人脸是否被较大比例的遮挡。为此我们需要训练能够输出人脸遮挡比例的模型,在做人脸矫正过程中计算人脸被遮挡的比例,对于人脸处于较大遮挡的情形,我们便不做矫正。因此,本文的两个工作重点是从单张图像中重建出高质量的、能够应用于人脸矫正等下游应用的三维人脸模型和训练能够输出人脸遮挡比例的模型。以往的三维人脸重建训练过程中使用的均为人脸68个关键点检测方法。但是这种方法存在关键点检测过少,检测的准确度不高以及不够鲁棒等问题。为了使得三维人脸重建模型更加稳定、精度更高,本文使用了检测人脸240关键点的算法,关键点的检测更加准确,能够适应不同的场景和光照。同时蕴含更多的人脸信息,重建出的三维人脸模型的质量也会更好。在以往的研究中,模型不能够很好地应用实际运用中的关键的一点在于训练数据不足、训练样本不够丰富。在本文中,为了得到高质量、能够应用于实际运用中的三维人脸模型,我们采集了大量的视频数据,建立了更加庞大的训练数据集。这个数据集总共约80万张人脸图像,包含不同的场景,不同的人脸姿态,不同的光照,不同的人种等情形。使得训练得到的人脸模型能够满足实际应用的需求。在以往的研究中,人们往往注重于被遮挡的人脸关键点检测、分割出人脸被遮挡的区域等任务,而很少去研究训练出能够输出人脸被遮挡的比例的模型。本文在原有的训练人脸重建网络的基础上加以改进,得到能够输出遮挡比例的模型。我们从80万张图片中选出约20万张图片(包含无遮挡),利用全卷积神经网络计算人脸被遮挡的比例,将其作为真值训练,得到的模型具有较高的精度。更多还原

【Abstract】 With the development of computer science,videoconferencing has become a common phenomenon.However,faces which are in videos seem to look down or look up in some facilities due to the angle of camera.Thus,correcting the angles of these faces became a new demand.A solution to this problem is that we can construct a 3D model from a single image and then rotate this model and project it to 2D plane.The basis of correcting angles of faces is getting 3D face reconstruction model which is robust and high quality.Meanwhile,we need compute proportion of occluded faces.If this proportion is large,we don not correct the angle of the face.Thus,we are required to do two works.One is training 3D face reconstruction mode from single image using large face pictures.Another is training a model that can output proportion of occluded faces.People are used to utilize the method that can detect 68 key points for 3D construction,It has some drawbacks.For instance,the number of the key points are kind of small and 3D model will not be high-quality.In order to get robust and high-quality 3D face reconstruction,we use the algorithm that can detect 240 facial feature points instead of the classic algorithm that can detect 68 facial feature points.In previous works,the lack of training data can lead to impractical 3D models.In order to solve the problem that the training data is not sufficient and various,we created a larger training data which can cover different scenes,illumination,poses,races.The number of face pictures is about 0.8 million.In early works,people prefer to research the problems like Occlusion key points detection and face segmentation.We are interested in computing the proportion of occluded faces.We trained a model that can show us the proportion of occluded faces.We use the method of supervised training based on the net that trains 3D face reconstruction model.We selected 200k pictures to compute the true proportion of occluded faces.更多还原