一种未知光源参数的明暗恢复形状方法

针对传统SFS(shape from shading)必须已知光源参数的缺陷，提出了一种新的使用神经网络恢复单幅未知光源参数环境中物体三维形状的方法。该算法利用前向神经网络的非线性映射能力，建立了物体表面形状和其对应的图像灰度值之间的非线性关系，所得权值可视为环境光源参数，由此得出反射图函数。基于该反射模型，物体表面高度值通过迭代的方法求得，并使用多分辨率分级实现SFS算法以减小算法复杂度。实验结果表明该算法对于无光源环境，能给出有效的恢复结果。相比传统算法，精度提高了约29％。

A New SFS(Shape from Shading) Algorithm with Light Source Unknown

Ref.5 proposed a SFS algorithm with light source unknown, provided that a training sample of an object whose shape is known under the same unknown light source is available. Like Ref.5, our algorithm uses forward neural network. Unlike Ref.5, our new SFS algorithm can do without the above-mentioned training sample. In the full paper, we explain in much detail our new SFS algorithm. Here we give only a briefing of our explanation. First we introduce briefly the traditional reflectance model. Taking advantage of the complex nonlinear reflectance mapping ability of the forward neural network, we establish a new reflectance model and reformulate the shape from shading problem as the minimization of an error function over the network weight. Object surface is recovered with recursive algorithm. And hierarchical implementation is used to reduce the computation cost. We performed simulations for two positions of unknown light source relative to object. In the case of a vase as object, our new SFS algorithm improved the recovering accuracy by an average of about 29% as compared with that obtained with the traditional method in Ref.4.