基于单幅图像的三维形状复原技术研究

发布时间：2019-03-14 12:18

【摘要】：基于二维图像的三维形状恢复技术是计算机视觉的前沿研究问题。明暗恢复形状是从单幅图像恢复物体三维形状的一种重要技术,其原理是建立三维曲面形状与图像灰度值之间的关系方程,经过求解方程来恢复三维曲面的深度信息。本文针对现有明暗恢复形状算法的不足之处,提出了基于改进Phong光照模型的优化算法,主要研究内容如下:(1)明暗恢复形状技术中的光照模型和投影模型的研究。针对传统明暗恢复形状方法在光照模型存在的局限性,本论文系统的研究了现有常用光照模型的适用范围和不足之处,在此基础上提出了一种改进的Phong光照模型。采用Oren-Nayar反射模型代替Phong模型中的漫反射部分,避免了原有模型中采用朗伯体模型描述物体表面漫反射所带来的误差。针对传统明暗恢复形状算法中采用正交投影模型这一弊端,改用符合实际照相机成像关系的透视投影模型,建立更有效的辐照度方程。(2)数值算法研究。研究了正交投影下基于朗伯体模型SFS算法、透视投影下基于朗伯体模型的SFS算法,分析了两种算法求解辐照度方程的原理过程、算法的精度和时间复杂度。针对这些数值算法求解精度不高的问题,本文将Phong模型下建立的辐照度方程转换成包含深度信息的哈密顿-雅可比偏微分方程,同时引入高阶局部LF通量分裂格式和改进的五阶WENO格式求解HJ方程,优化了HJ方程的求解精度,进而有效地提高了明暗恢复形状算法的重构精度。为了验证本文所提基于改进Phong模型的优化SFS算法的有效性。分别以单幅合成图像和真实图像为例进行三维恢复,并将图像的重建结果与现有算法的恢复结果进行定量和定性对比。经过实验验证,证实了本文所提算法在视觉效果上有明显提高。与基于正交投影的算法相比,本文平均误差下降了62.2%;与同类算法相比,本文平均误差下降了30.7%。
[Abstract]:Three-dimensional shape restoration based on two-dimensional image is a cutting-edge problem in computer vision. Shape restoration is an important technique to recover 3D shape from a single image. Its principle is to establish the relationship equation between the shape of 3D surface and the gray value of the image, and to recover the depth information of 3D surface by solving the equation. Aiming at the shortcomings of the existing shape restoration algorithms, this paper proposes an optimization algorithm based on the improved Phong illumination model. The main contents are as follows: (1) the illumination model and projection model in the shape restoration technology are studied. In view of the limitations of traditional shading restoration methods in illumination models, this paper systematically studies the scope of application and shortcomings of the existing lighting models, and proposes an improved Phong illumination model based on this. The Oren-Nayar reflection model is used to replace the diffuse reflection part of the Phong model, which avoids the error caused by the Lambert body model used in the original model to describe the diffuse reflection on the surface of the object. Aiming at the disadvantage of orthogonal projection model used in traditional shape restoration algorithm, a more effective irradiance equation is established by using perspective projection model which accords with the actual camera imaging relationship. (2) the numerical algorithm is studied. The SFS algorithm based on Lambert body model under orthogonal projection and SFS algorithm based on Lambert body model under perspective projection are studied. The principle and process of solving irradiance equation by two algorithms are analyzed. The accuracy and time complexity of the algorithm are analyzed. In this paper, the irradiance equation based on Phong model is transformed into Hamiltonian-Jacobian partial differential equation with depth information. At the same time, the high-order local LF flux splitting scheme and the improved fifth-order WENO scheme are introduced to solve the HJ equation. The accuracy of the HJ equation is optimized and the reconstruction accuracy of the shape restoration algorithm is improved effectively. In order to verify the effectiveness of the proposed optimization SFS algorithm based on the improved Phong model. A single composite image and a real image are taken as examples for three-dimensional restoration, and the reconstruction results are compared quantitatively and qualitatively with those obtained by the existing algorithms. The experimental results show that the proposed algorithm has a significant improvement in visual effect. Compared with the algorithm based on orthogonal projection, the average error of this paper is reduced by 62.2%, and the average error of this paper is reduced by 30.7% compared with the same algorithm.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】