面向各类反射表面的光度立体视觉方法研究

发布时间：2018-11-20 04:23

【摘要】：自从Woodham首次提出光度立体视觉后,该技术在表面三维重建中得到了广泛应用。在光度立体视觉中,物体表面由多个方向光依次照射并由一个视线方向固定的相机采集。为了更好地处理具有各类双向反射分布函数(BRDF)的材质,现有研究提出了基于朗伯漫反射假设和基于BRDF的光度立体方法。近些年来,一些方法具有了较好的性能,但在部分应用过程中还需进一步解决以下问题：(1)提高法向重建精度；(2)合理减少光源数量；(3)提升算法效率。本论文针对上述问题开展了以下创新性研究工作：1.提出了共线光源光度立体的改进方法。该方法将高光检测转化为多分类器和多训练集的模式识别问题,提升了高光检测的精度,并在法向重建中使用了l1范数的正则项,使法向估计更为鲁棒。2.提出了基于反射分量一般特性的光度立体方法。该方法对漫反射的平滑性、高光的集中分布性、阴影的低强度性质进行统一建模,并通过一个图模型将这三项分别表达为反射率局部变化总和、分组稀疏以及加权的l1范数项。该优化问题可通过二次锥优化求解,对各向同性和各向异性材质均有很好的效果。3.提出了基于核回归的光度立体方法。该方法使用可变参数的核回归对反射进行了隐式建模并将其转化为特征值求解问题,其主要计算开销为Gram矩阵求逆以及3×3矩阵的特征向量求解。最优参数通过留一法交叉验证确定,并通过适当处理对该过程进行加速。该方法在法向估计精度和计算效率上都取得了较好的性能。值得指出的是,与现有基于BRDF的光度立体方法相比,后两项工作对反射率表示进行简化,将其视为仅含单个变量(光源方向)的函数,从而有效地降低了反射率函数的维度。正是这方面的创新使得这两个方法在各类反射材质上都取得了很好的性能。
[Abstract]:Since Woodham first proposed photometric stereo vision, this technique has been widely used in surface 3D reconstruction. In photometric stereoscopic vision, the object surface is illuminated by multiple directional light in turn and captured by a camera with a fixed line of sight. In order to better deal with materials with various bidirectional reflectance distribution functions (BRDF), a photometric stereoscopic method based on Lambert diffuse reflection hypothesis and BRDF is proposed. In recent years, some methods have better performance, but the following problems need to be solved in some applications: (1) improving the accuracy of normal reconstruction; (2) reducing the number of light sources reasonably; (3) improving the efficiency of the algorithm. In this paper, the following innovative research work has been carried out in view of the above problems: 1. The improved method of collinear luminosity stereo is put forward. This method transforms the specular detection into multi-classifiers and multi-training sets, improves the accuracy of the specular detection, and uses the regular term of L _ 1-norm in the normal reconstruction, which makes the normal estimation more robust. 2. A photometric stereoscopic method based on the general characteristics of reflectance component is proposed. In this method, the smoothness of diffuse reflection, the concentrated distribution of specular light and the low intensity property of shadow are modeled uniformly, and the three terms are represented by a graph model as the sum of local variations of reflectivity, group sparsity and weighted L 1 norm term respectively. This optimization problem can be solved by means of quadratic cone optimization, and it has a good effect on both isotropic and anisotropic materials. A stereoscopic photometric method based on nuclear regression is proposed. The method uses variable parameter kernel regression to model reflection implicitly and transforms it into eigenvalue solving problem. The main computational overhead is the inverse of Gram matrix and the solution of eigenvector of 3 脳 3 matrix. The optimal parameters are determined by a method of cross-validation, and the process is accelerated by proper processing. The method achieves good performance in normal estimation accuracy and computational efficiency. It is worth pointing out that compared with the existing photometric stereoscopic methods based on BRDF, the latter two works simplify the reflectivity representation and treat it as a function containing only a single variable (the direction of the light source), thus effectively reducing the dimension of the reflectivity function. It is the innovation in this field that makes the two methods achieve good performance in all kinds of reflective materials.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】