GPU加速的鲁棒性人脸2.5D重建方法

针对人脸三维重建问题，提出利用多面体模型快速从人脸三幅光度立体图像中重建其2.5D结构的鲁棒性方法，并较少受高光和阴影以及毛发的影响。以漫反射光照模型建立多面体每个面与人脸图像每个像素点间的方程组，利用加权最小二乘方法求解初始人脸2.5D结构，并利用计算机图形学算法以包含镜面反射和阴影的绘制模式虚拟合成该初始结构的光度立体图像。以此虚拟光度立体图像进行第二次重建，并比较两次重建结果的差异以推断重建方法的误差，从而修正计算结果。为加速计算过程，给出了充分利用现代可编程图形处理器(GPU)高速并行计算能力的硬件实现计算的方法。在YaleB人脸数据库上的实验表明，该方法能快速重建出人脸的2.5D形状和纹理，并克服阴影和高光以及毛发的影响。

A Robust Method for 2.5D Face Reconstruction Accelerated by GPU

A robust method for 2.5D facial shape and texture reconstruction based-on the model of polyhedron and the technology of photometric stereo was presented. Using the continuity restrictions and error-reduced operator, the presented method can overcome the influence of shadow, highlight and hair. The 2.5D shape and texture of human face was reconstruction firstly by solving a linear equations system, which was established by the corresponding relationships between faces of polyhedron and pixels of human facial images under three independent illuminations. Then, three virtual photometric stereo images were synthesized from reconstructed results using computer graphics. The local shadow analysis and specular reflection are used to make the synthesized images approach to the real images. Facial shape and texture were reconstructed again from synthesized virtual photometric stereo images. Finally, these two reconstruction results are combined to reduce the reconstruction error and produce a superior result. A fragment shader based-on GPU was presented to accelerate computation process. Experimental results on YaleB database show that this method can reconstruct the 2.5D shape and texture of human face rapidly and robustly.