CASIA OpenIR  > 毕业生  > 博士学位论文
基于静态图像的人脸三维重建和识别相关问题研究
其他题名Research on Issues of Face 3D Reconstruction and Recognition Based on Static Images
熊鹏飞
学位类型工学博士
导师刘昌平
2012-05-27
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词人脸识别 人脸三维重建 主动形状模型 眼镜遮挡 光照 深度图像融合 异质图像 Face Recognition Face 3d Reconstruction Asm Glasses-occluded Illumination Heterogeneous Binocular Stereo Image Fusion
摘要人脸三维重建与识别一直是计算机视觉、计算机图形学与模式识别领域中的热点研究内容。作为人脸视觉感知技术的一部分,它在身份认证、视频监控、人机交互、娱乐动画等领域有着广泛的应用前景。现有的人脸重建与识别方法在受控条件下均能取得较好的性能,但受限于人脸图像平滑,以及遮挡、姿态、光照、异质模态等采集环境变化的影响,真实感人脸模型重建与人脸识别都受到了很大的挑战。 在对前人工作进行分析总结的基础上,本文深入研究了基于静态图像的人脸三维重建技术,并结合三维模型,针对人脸图像识别问题提出了应对不同干扰的解决方法。该研究课题基于这样的出发点,不同的应用场合对人脸模型的精度和平滑度的要求不同,需要研究多种三维重建方法;恢复人脸三维形状,可以有效补偿人脸成像中的特征丢失,增强环境变化下人脸识别的鲁棒性。具体而言,本文的主要工作和贡献可归纳如下: 1、基于标准模型形变的人脸三维重建。(1)针对传统的主动形状模型(ASM)中初始化敏感与姿态鲁棒性低的问题,提出了一种基于关键点定位与三维形状约束的多层ASM形状定位策略。该方法有效利用了人脸形状中器官关键点定位精度最高,而轮廓点最低的特性,通过关键点定位结果构建径向基函数变换模型(RBF)来实现其它形状点的初始化,提高了整体形状的初始化精度;进一步,通过估计人脸三维姿态获得轮廓点的平面偏移,避免了初始化与迭代中原始二维仿射变换对轮廓点的估计不足,同时提高了ASM的定位精度与姿态鲁棒性;(2)针对直接基于二维形状点的模型形变容易在人脸姿态变化时出现模型扭曲的问题,采用前述姿态估计方法基于原始人脸进行迭代的模型变形和姿态矫正,提高了重建模型的平滑性与真实度。 2、基于双目立体视觉的人脸三维重建。(1)针对人脸图像过于平滑而难以获得稳定的关键点匹配的问题,提出了基于LOG算子的关键点检测与RBF的匹配扩展方法。该方法有效提高了传统SIFT算子中DOG检测子所获得的角点密度,进而通过建立已匹配点对与未匹配点之间的映射模型来获得匹配点扩展,改善了传统三角面片扩展中的区域误匹配性,实现了分布合理且数目充分的关键点匹配;(2)针对全局匹配模型噪声较多而局部匹配过于平滑的特性,提出了全局匹配与局部匹配结合的稠密匹配方法。先通过种子点繁殖的双向匹配来获得稳定的匹配点,进而对匹配点之间的空洞采用边缘的动态规划来进行填补,恢复了足够平滑且准确的人脸三维形状。 3、基于三维模型的眼镜遮挡下人脸识别。针对现有解决眼镜遮挡的算法容易造成图像纹理特征失真的问题,提出了基于三维人脸模型生成眼镜遮挡下人脸虚拟样本来扩充样本库的方法。该方法将眼镜遮挡下图像看作整体,通过人脸三维重建来还原眼镜对图像的遮挡,有效消除了眼镜多样性带来的人脸纹理不稳定。进一步,针对眼镜镜片固有的反光与模糊现象,分别提出了束光源与混合纹理的方法来进行模拟,提高了虚拟样本的真实感。虚拟样本的思路和镜片的细节处理均提高了人脸识别性能。 4、基于三维人脸对齐的异源模态下人脸识别。针对人脸识别中的光照与异质模态,分别提出了三维人脸对齐的同...
其他摘要Both face 3D reconstruction and recognition based on images are the significant researchs in the fileds of Computer Vision, Computer Graphics, and Pattern Recognition. As the primary components of Face Perception, these two technologies have broad application prospects in identity authentication, video surveillance, face animation, human computer interaction, and so on. Currently, the start-of-art face reconstruction and recognition system achieves satisfied results under the well-controlled environment. However, they are still challenged by the facial smooth texture and various image environments, such as occluded, posture, illumination or heterogeneous capturing conditions. With a thorough review of previous works, this paper researchs the issues of face 3D reconstruction and its supplements on face image recognition, which is based on the following starting points, that the accuracy and efficiency of facial shape reconstruction is different under the various applications, which leads to the employment of kinds of face 3D reconstruction methods, also facial shape can effectively compensate the texture loss in face imaging, while face is known as 3D nonrigid object. In particular, the main contributions of this thesis include following parts: 1) Face 3D reconstruction based on deformable model. Firstly, to avoid the drawbacks of initialization sensitivity and low location precision of profile face contour, a multi-layer ASM based on keypoints location and 3D shape adjustment is proposed as the basis of face 3D reconstruction, while the location precision of keypoints is better than others. In shape initialization, keypoints are located firstly to fit a RBF model for the transformation of other points. Then the 3D facial posture is estimated to take the place of the original 2D shape adjustment and compensate the misregistration of contour points. Both of them improve the traditional ASM location precision and stability. Secondly, to handle with the model distortion in the directly transformation based on 2D shape points, the original image is aligned based on the 3D estimation posture, which mproves the flatness of the releastic model. 2) Face 3D reconstruction based on binocular stereo vision. Firstly, a keypoints matching approach based on LOG detector and RBF extension are carried out to generate denser and acattered matching pairs on the sparse facial texture, which enhance the points density in the regular SIFT detector. Also the RBF mapping bet...
馆藏号XWLW1749
其他标识符200918014628060
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/6426
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
熊鹏飞. 基于静态图像的人脸三维重建和识别相关问题研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2012.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA_20091801462806(5166KB) 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[熊鹏飞]的文章
百度学术
百度学术中相似的文章
[熊鹏飞]的文章
必应学术
必应学术中相似的文章
[熊鹏飞]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。