【作者】 程岳；

【导师】 沈会良；

【作者基本信息】 浙江大学 ， 物理电子学， 2013， 博士

【摘要】 场景表面的三维立体重建是计算机视觉领域的一个重要分支,其在产品质量控制、工业测量、医学诊断、文物的数字化保存、现场取证、数字娱乐以及虚拟现实等多种领域都有着广泛的应用。在一些应用领域,基于摄像机多视角几何的方法往往不能精确重建场景表面的局部细节特征,使得最后输出三维模型的真实感大大降低,不能够满足实际应用要求。而基于光度立体的测量方案恰好能够较好地表现重建物体的局部细节,极大地补充和完善了基于多视角几何的方法,在研究领域得到了广泛的关注。自从Woodham第一次提出光度立体视觉的概念以及实现方法以来,由于其在三维场景细节重建方面的巨大潜力,三十年来一直保持着持续的关注度,相关学者在这些方面做出了许多有意义的研究与探索。本文在深入分析了近年来国际上光度立体视觉领域相关问题的基础上,重点研究了光度立体视觉中的光源参数标定问题、非朗伯表面法向估计问题以及基于梯度场的表面重建问题,通过对光度立体视觉中这三个关键步骤高精度算法的实现,保证了最后高质量的光度立体三维重建结果。本文的主要工作和创新之处在于：(1)参考传统摄像机标定算法,利用镜面反射和理想漫反射中的几何光度信息,提出了一种基于平面靶标的光源标定算法。区别于传统基于球面的光源标定算法,算法对标定靶的要求比较低,仅仅利用常见的平面镜,划分出镜面反射与漫反射区域,就可以方便地实现光源标定。本文提出的光源标定算法精度相对于传统标定算法提高了一个数量级,从而能够保证后续光度立体视觉法向估计的精度。(2)基于线性光源系统,利用光度图像之间的线性偏差,提出了一种非朗伯表面的法向估计算法。算法在线性光源结构下,首先把高光检测问题转化成线性偏差的模式分类问题。利用经过实际数据训练或者模拟数据训练的分类器,可以较好地把高光像素从漫反射像素中进行检测,使得光度图像中剩余的少量高光误差呈现较为稀疏的分布。最后利用基于e1范数逼近的法向估计算法,能够较好地校正稀疏分布的高光误差,使得法向估计精度进一步提高。(3)借鉴模式识别中的核方法,提出了一种基于核化泊松方程梯度重建算法。算法对传统的基于泊松方程的梯度重建算法进行核化改造,使得重建过程中每个像素的对应曲面高度不仅与该点观测到的梯度相关,并且与多个邻域像素相关。通过对重建过程中边界问题的仔细考虑,保证重建曲面的边界平滑,使得从具有高斯分布噪声的梯度场数据能够较好地重建目标曲面。(4)参考压缩感知领域稀疏表达的思想,提出一种基于e1线性解码的梯度重建算法。对于呈稀疏分布的梯度噪声,算法利用e1线性解码方法从受到污染的梯度数据到目标曲面的高度数据进行重建,并在解码框架中加入拉普拉斯正则项,极大地增加了算法在稀疏噪声下的解码能力。实验证明,在多达30%以上像素点存在稀疏噪声的情况下,算法依然能够稳定地从梯度数据进行表面重建。更多还原

【Abstract】 Three-dimensional surface reconstruction is an important branch of Computer Vision. It’s widely used in the quality control of products, industrial measurement, medical diagnosis, digital preservation of relic, scene evidence, digital entertainment and virtual reality. In some application areas, local features of measured object can not be well reconstructed based on the geometric measurement methods which makes the details of reconstructed surface quality greatly degraded, and often can not meet the requirements. On the other hand, Photometric Stereo scheme can keep local details of measured objects well. It’s quite satisfactory for Computer Vision applications and greatly complement geometric measurement based methods. Thus Photometric Stereo has received wide attention in the research field.Since Woodham first proposed the concept and realization method of Photometric Stereo, its huge potential in the three-dimensional reconstruction of scene details has been noticed. In the past thirty years, scholars have done many meaningful research and exploration in this area and it is still a hot topic and trend now. After analyzing the recent research approach in Photometric Stereo field, we have focused on the light sources calibration, non-Lambertian surface normal estimation and shape from gradient fields. By keeping highly precision in each step of Photometric Stereo using proposed methods, the final reconstruction outperform state-of-the-art techniques. The major work and contributions lie in several fields as follows:(1) Using the parameters acquired by camera calibration and the geometric&photometric in-formation from the planar surface, we propose a novel method for light sources calibration. Differ-ing from the traditional sphere based methods, proposed method just uses a plane mirror which is divided into mirror reflection area and diffuse area respectively. The accuracy of proposed method is improved by an order of magnitude compared to the traditional methods. So it can guarantee the precision of normal estimation in Photometric Stereo.(2) Based on co-linear light source configuration, we propose a non-Lambertian surface nor-mal estimation method. Using deviations in photometric images under the co-linear light sources, the specularity detection problem is converted to pattern classification one. In training step we use real or synthetical data to build a robust specularity classifier. After discarding specularity in testing step, the residual error in photometric images is sparsely distributed. An l1-norm approximation method is designed to further correct the sparse error. The final output is quite satisfactory.(3) We convert the traditional Poisson equation to its kernel representation form using kernel method, the underlying surface can be recovered from Gaussian noise contaminated gradient data by kernel regression. With further considering an even extension of the original gradient field, we assume a Neumann boundary condition which makes final reconstructed surface smooth and reliable.(4) Inspired by the idea of sparse expression in compressive sensing fields, we propose andecoding method for surface reconstruction from gradient fields. The l1decoding procedure can exactly recover surface from sparse noise contaminated gradient data. The Laplacian term is additionally employed to increase the information in decoding matrix and suppress noise and/or outliers. Experimental results validate that the proposed method significantly outperforms state-of-the-art techniques, and can produce satisfactory reconstruction even in the very extreme situation of60%outliers.更多还原