【作者】 阳瑜；

【导师】 吴小俊；

【作者基本信息】 江南大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 随着计算机视觉技术的发展,我们对计算机的图像处理分析能力的需求也越来越高。在模式识别中,人脸识别有着极其重要的地位和应用价值。传统的人脸分析、识别技术都是基于2D人脸图片的,但是自然条件下,由于姿态、表情、光照等因素的变化,同一个人的2D图片在这变化的因素下存在非常大的差异。为了解决这些刚性和非刚性变化给人脸识别、分析带来的困难,人脸的特征表示、特征提取方法被提出用来表示人脸内在生物识别信息的不变性并且达到人脸图片降维的目的。相对于2D人脸,3D人脸具有内在的姿态和表情不变性,而且3D人脸模型可以用来对2D人脸进行光照情况的分析以及估计相应的光照参数。随着3D人脸数据采集设备的精确度提升以及3D人脸重构算法的发展,3D人脸的相关研究近年来收到了研究者们的广泛关注。对比2D人脸,3D人脸相包含更多的信息,在生物识别上具有更好的发展前景。本文在3D人脸形状重构,3D人脸模型在2D人脸图片标准化中的应用以及3D人脸模型纹理重构这几个方面做了一些研究。从现在广泛使用的3D人脸模型和3D人脸重构算法出发,本文的主要贡献分为以下几个方面:(1)理论上详细介绍了传统3D人脸重构模型的算法基本思想和人脸重构原理。并且在不同的人脸数据库上进行人脸重构实验,直观的分析传统方法在3D人脸重构上的优缺点。(2)提出了人脸特征点加权的3D人脸形状重构算法。现今广泛使用的3D人脸重构模型3D Morphable Model(3DMM)能从一张2D人脸图片重构出相应的3D人脸,随着人脸特征点定位算法的发展,经典的3D人脸重构算法利用2D人脸特征点进行3D人脸重构,但是改进后的重构算法在3D形状估计时没有考虑人脸特征点的语义信息和不同特征点对人脸重构的不同影响。因此,我们提出特征点加权的3D人脸形状重构算法。我们引入特征点权重信息对不同的特征点进行加权处理,使得我们的算法在3D人脸形状重构过程中可以对单独的人脸特征点根据其重构精度动态调整权重,提升了这些具有不同语义信息的特征点的重构精度,从而进一步提高了3D模型整体的重构精度。我们在不同的人脸数据库上进行了3D人脸重构实验,分析该算法在特征点级别和3D人脸的整体误差提升,并且在不同人脸表情和姿态条件下验证了算法的有效性。(3)我们改进了现有最新的基于3DMM的2D人脸图片的标准化算法。2D人脸标准化在人脸识别中是一个重要的步骤,最开始的人脸标准化方法有人脸定位检测,人脸灰度归一化,直方图均衡化等,但是这些传统人脸图片处理方法无法对2D人脸图片的姿态和表情进行归一化。基于3DMM的2D人脸图片标准化方法从2D图片中估计对应人脸的姿态和表情信息并且重构相应的3D人脸模型,然后根据模型的参数信息对2D图片中人脸的姿态和表情进行相应的调整来达到人脸标准化的效果。但是现有的基于3D模型的人脸标准化算法对人脸自遮挡区域重构的效果并不好。重构的自遮挡填充区域不光滑、视觉效果不自然,和周围区域相比有纹理差异。为了解决这个问题我们提出了新的自遮挡区域的纹理重构方法。改进了自遮挡区域的纹理重构算法后,实验得到的填充区域具有真实的光照强度和人脸纹理细节,人脸标准化的图片更加自然。(4)提出了基于人脸标准化的3D人脸重构算法。现有的3D人脸重构算法都是基于3D模型来估计形状参数和纹理参数,根据这些3D参数对3D模型进行形变和纹理渲染,得到最终的重构结果。但是这些方法的人脸纹理重构是通过建立光照模型来估计光照参数,进而根据光照信息改变模型的平均纹理来得到纹理重构结果。由于传统方法只用少数的纹理参数表示真实人脸纹理,而且人脸模型的平均纹理表达能力也非常有限,所以得到的人脸的纹理会丢失很多人脸细节纹理信息,比如:胡须、皱纹、瞳孔颜色、胡须颜色等。这些人脸细节纹理信息不仅很重要,而且具有很重要的生物鉴别信息。为了重构得到更为精确自然的3D人脸纹理,我们结合了基于3DMM的人脸正则化方法,从标准化之后得到的正面人脸图片中提取人脸纹理信息,由于人脸正则化过程估计得到了相应的人脸形状参数,我们再结合提取的纹理信息通过渲染即可得到最终的3D人脸重构结果。相较于现有的3D人脸重构算法,我们提出的方法能得到更为精确细致的3D人脸重构结果。我们在不同的2D人脸数据库上做3D人脸重构实验。实验结果表明,我们提出的方法在人脸细节纹理特征的重构结果有明显的提升。更多还原

【Abstract】 With the development of computer vision technology,we are increasingly demanding the image processing and analysis capabilities of computers.In pattern recognition,face recognition has an extremely important position and application value.Traditional face analysis and recognition techniques are based on 2D face images,but under the natural conditions,the 2D facial image of the same person have large difference due to changes in pose,expression,illumination and other factors.In order to solve the challenge in face recognition and analysis caused by the facial rigid and nonrigid variation,the facial representation and feature extraction methods are proposed to represent the invariance of the biometric information in the face and achieve the purpose of reducing the dimension of the face image.Compared with the 2D face,the 3D face has an intrinsic pose and expression invariance.Moreover,the 3D face models can be used to analyse and estimate illumination condition for 2D face images.With the improvement of the accuracy of 3D face data acquisition devices and the development of 3D face reconstruction algorithms,the research on 3D face has received extensive attention from researchers in recent years.Compared with 2D faces,3D faces contain more information and have a better development prospect in biometrics.In this paper,we have done some researches on 3D face shape reconstruction,3D face model application in 2D face image normalization and 3D face texture reconstruction.Starting from the widely used 3D face model and 3D face reconstruction algorithm,the main contributions of this paper are divided into the following aspects:(1)In theory,the basic idea of the traditional 3D face reconstruction model and the principleof face reconstruction are introduced in detail.We performed 3D face reconstructionexperiments on different face databases to intuitively analyze the advantages anddisadvantages of traditional methods.(2)A new 3D face reconstruction algorithm based on face feature points weighting isproposed.The 3D Morphable Model(3DMM),which is widely used today,canreconstruct the corresponding 3D face from a 2D face image.With the development ofthe face feature point localization algorithm,the traditional 3D face reconstructionmethod uses 2D feature points for 3D face reconstruction,which improves theefficiency of the algorithm.However,the improved reconstruction algorithm does notconsider the semantic information of face feature points and the different influences ofdifferent feature points on 3D face shape reconstruction.Therefore,we proposed thefeature point weighted 3D face shape reconstruction algorithm.We introduce featureweight information to weight different feature points.This allows our algorithm todynamically adjust the weight of individual face feature points according to itsreconstruction accuracy in the 3D faces shape reconstruction process.Considering theweight of each feature points individually enhances the reconstruction accuracy offeature points which represent different semantic,thereby further improving the overallreconstruction accuracy of the all 3D model.We performed 3D face reconstructionexperiments on different face databases,analyzed the overall error of the algorithm atfeature point level and 3D face.And verify the effectiveness of the proposed algorithmin different facial expressions and poses.(3)We have improved the existing normalization algorithm for 3DMM-based 2D faceimages.The 2D face normalization is an important step in face recognition.The initialface normalization methods include face detection and localization,face grayscale,histogram equalization,etc.but these traditional face image processing methods cannotnormalize the pose and expression of 2D face images.The 2D face image normalizationmethod based on 3DMM estimates the shape,pose and expression parameters fromimage.Then,the corresponding 3D face shape is obtained according to these 3Dparameters.Finally,according to the pose and expression regularization of the 3D shape,the posed and expression of the 2D face are adjusted accordingly to obtain a normalized2D face image.However,the existing 3DMM-based face normalization algorithms arenot effective in reconstructing the face self-occlusion region.The reconstructed self-occlusion fill area is unsmooth and unnatural in visual,the texture of the filling area hasdifferent texture compared with the surrounding area.In order to deal with this problem,we a new texture reconstruction method for self-occlusion regions.With theimprovement of algorithm,the reconstructed filled area has real illumination intensityand face texture details and the face normalized image is more natural.(4)A 3D face reconstruction algorithm based on face normalization is proposed.Theexisting 3D face reconstruction algorithms estimate the 3D shape and texture based on3D model.In the fitting process,the model is deformed and texture rendered accordingto the 3D parameters and the 3D face reconstruction result is finally obtained.However,the face texture reconstruction of these methods estimates the texture by establishingillumination parameters by establishing the light model,and then change the averagetexture of the model according to the illumination information to obtain thereconstructed texture.Since the traditional method only uses a few texture parametersto represent the real face texture and the average texture representation ability of theface model is very limited,the fitting result of face texture loses a lot of face detail information,such as: beard,wrinkles,pupil color,beard color,etc.But these details textures have important biometric information.In order to reconstruct the more accurate and natural 3D face texture,we combine the 3DMM-base face normalization method to extract the face texture information from the frontal view face images.The face normalization process estimates the 3D shape parameters for the corresponding 2D face image also,so we can get the final 3D face reconstruction result by rendering the extracted texture onto 3D shape.Compared with the existing 3D face reconstruction algorithms,our proposed method can get more accurate and detailed 3D face results.We have done the 3D face reconstruction experiments on different face database.The experiment results show that the proposed method has a significant improvement in the reconstruction of facial detail texture features.更多还原