基于图像对的立体重建是用于获取人脸三维信息的通用方法, 但根据图像数据和重建算法所得到的三维重建结果存在各种误差, 本文对通用形变模型进行改进并与三维立体重建融合以得到更精确的重建结果。首先使用Max-Margin对象检测算法来获取面部边界框, 其中回归树集合法能直接从像素强度的稀疏子集识别面部特征点。然后通过PCA颜色模型生成形状和颜色的三维面部统计模型, 利用ISOMAP算法将三维网格转换为二维表面并提取纹理信息, 得到面部模型。最后在源网格上进行两步非刚性表面配准的变形过程: 先通过对源网格进行二次采样来选择少量网格点来表示源的全局变化, 并选取径向基函数(RBF)进行非刚性全局变形; 再对源顶点进行Procrustes分析获得非刚性变换, 再通过加权方案来进行k-近邻变换, 得到平滑的局部变形。将单图像重建的面部模型, 立体重建的面部模型和本文的面部变形模型与高质量扫描云图进行对齐比较, 得到面部变形模型的3个RMS值分别为2.795 2, 2.102 8和2.153 4, 相比于其他模型, 面部变形模型更接近高质量扫描云图, 即与原图像一致性更高, 误差更小。面部变形模型的定性和定量分析表明, 立体重建与人脸一般形状信息的组合在几何信息的表达上优于基于通用模型的单个图像重建以及未考虑通用模型的立体重建。

Stereo reconstruction from image pairs is a genetic method for 3D acquisition of human faces. Depending on available imagery and reconstruction algorithm, the resulting 3D reconstructions may have deficits. We improve generic morphable model and fused with stereo reconstruction to remedy such deficits. Firstly, we obtain the face bounding box by Max-Margin Object Detection, and recognition the face feature points directly from a sparse subset of pixel intensities by the Ensemble of Regression Treesmethod. Secondly, through the color PCA model to generate the shape and color of the 3D facial statistical model, the ISOMAP algorithm is used to convert the 3D mesh to 2D surface and extract the texture information to get the facial model.Finally, the surface registration with two step non-rigid change on source mesh: we mean the global deformation in the source by using few anchor points on the source mesh, and using the Radial Basis Function (RBF) to express non-rigid global deformation; and then Procrustes analysis is applied for non-rigid transformation of the source vertex, the k-nearest neighbor transformation through weighted scheme tosmooth the local deformation. The experiment result shows the high-quality scan which is used for cloud comparison with the face model, stereo reconstruction and our deformed face model, after aligning all models with the high quality scan, three RMS facial deformation model values are 2.795 2, 2.102 8, 2.153 4.Comparing to other models, deformed face model has smaller discrepancies dominate more strongly, and the deformed face model is visually more shape consistent to the image. Qualitative and quantitative analysis of the deformed face model shows that the combination of stereo reconstruction with general shape information about human faces is geometrically superior to both the reconstruction based on single image from generic model as well as the stereo reconstruction without consideration of the generic model.