【作者】 肖甫；

【导师】 吴慧中；

【作者基本信息】 南京理工大学 ， 计算机应用， 2007， 博士

【摘要】 可交互性、沉浸感和想象力一直是虚拟现实系统(VR系统)所追求的目标，而真实感和实时性是VR系统构建中的一对矛盾。传统基于几何的虚拟场景绘制技术是以计算机图形学为基础，可以准确获知场景的深度信息，因而用户与场景对象之间具有良好的交互性，且视点可以自由变换。但该方法的复杂度完全依赖于场景，限制了其在普通硬件平台上应用。基于图像的绘制方法将图像作为信息载体，通过对原图像的重采样、重组织的绘制过程，因而具有绘制速度独立于场景复杂度、真实感强等特点，并成为近年来虚拟现实技术的研究热点。本文以基于图像绘制的虚拟场景生成为研究主线，主要研究了图像拼接技术、基于图像的重光照技术、图像分辨率增强技术等关键技术。主要研究内容如下：图像拼接是虚拟场景建模的主要技术之一，针对光照变化图像序列提出了两种基于特征的图像拼接算法：基于静态小波序贯特征匹配和同态滤波的图像拼接算法以及基于圆投影和能量函数优化的图像拼接算法。其中基于静态小波序贯特征匹配和同态滤波的图像拼接算法采用静态小波提取图像匹配模板的特征序列，较好地实现了光照不同的图像匹配；而对于初始图像获取时光照条件的不同，采用同态滤波等技术解决初始图像的光照不一致问题。基于圆投影和能量函数优化的图像拼接算法则采用圆投影提取图像匹配模板的特征序列，克服了传统图像特征提取方法中的区域局限问题；对于图像融合，论文从图像能量角度出发，构造一个包含图像梯度的能量泛函，通过最小化该函数来求取图像重叠区域的全局最优融合因子，最终实现快速、准确地拼接图像。研究了光照条件对于场景图像的重要意义，提出了一种基于图像的重光照算法。该算法在给定目标若干基图像及其对应光照属性的基础上，通过基图像数据矩阵的奇异值分解(SVD)分离场景的环境光分量、反演场景的表面反射率和表面法向量等。新光照条件下场景的绘制，直接光照采用光照模型直接计算；间接光照采用分离出的基图像环境光分量拟合，较好地实现了变化光照下场景图像的生成。针对高分辨率真实感虚拟环境及场景浏览时变焦观察的需要，研究了图像分辨率增强算法。对于单帧图像的分辨率增强，提出一种基于熵变分的图像分辨率增强算法。该算法在贝叶斯估计和最大熵原理的基础上，将图像象素点梯度信息应用到图像分辨率增强中，从而建立起一种基于图像梯度信息的各向异性自适应分辨率增强算法。对于多帧图像序列的超分辨率复原，在单帧熵变分模型的基础上，将双边滤波技术引入到图像超分辨率复原中，建立了一种基于广义熵变分的图像超分辨率复原模型，从而提出一种基于几何距离和梯度信息的双重加权各向异性分辨率增强算法。实验表明：使用本文算法得到的超分辨率复原图像具有较高的峰值信噪比和视觉质量。针对基于图像的虚拟场景虚化关联的需要，将非真实感绘制技术引入虚拟场景中，提出了一种基于线积分卷积的场景虚化关联生成方法。该方法通过计算结构张量获得图像曲线走向矢量场，以此刻画图像局部结构；并采用全变分偏微分方程对曲线走向矢量场自适应滤波平滑处理；在此基础上利用线积分卷积指导画笔走向模拟生成线条波动感虚化效果图像；最后结合原始真实图像和虚化效果图像，实现了一种基于时间变量的柱面场景虚化关联算法，并为虚拟现实系统超越现实时空的需要提供了一种新思路。针对基于图像的虚拟全景空间漫游的需要，在定义位置链、时间链、焦距链和虚化链等链表结构的基础上，提出了一种包含位置关联、时间关联、焦距关联和虚化关联的大规模虚拟全景空间漫游策略。更多还原

【Abstract】 Interactive, immersion, and imagination are the objectives for virtual reality system, while reality and realtime are contradicted during VR system construction. Traditional virtual scene coustruction method based on geometry is on the basis of computer graphics; the depth information of scene can be obtained directly, which offers a convenient interactive mode between user and the scene. Also, user can change their viewpoint freely. However complexity of this method is entirely depends on virtual scene, and it is difficult to realize on common harder platform. A completely brandnew method-image based rendering is proposed recently, which uses image as the carrier of information. This method uses pixel as rendering element, rendering is just the process of resample and reorganize former pixel. Advantages of this method are rendering speed independent of scene complexity and high fidelity. So it becomes a hot reseach spot in virtual reality field.This dissertation concentrates on the key technology of virtual scene rendering based on image. Aimed at the needs of virtual panoramic space construction, we emphatic research the image stitching method; image based relighting method and image resolution enhancement method. Generally, the main research points in this paper are as follows:Image Stitching is one of the key technology in virtual reality construction,we proposes two ambient light independent panorama mosaics methods based on feature: image stitching algorithm based on stationary wavelet decomposition and homomorphic filtering, and image stitching algorithm based on ring projection and energy function optimization. The first algorithm uses stationary wavelet decomposition to extract feature sequences in image templates. Then color transform or homomorphic filtering can be used to adjust the global brightness of images to solve the problem of difference illumination, thus a pleasure stitching result can be achived. The second algorithm uses ring projection to extract feature sequences in image templates, which overpasses the local area limit in tradition methods. While blending factors are gained from energy function that contains image gradient, which conquer the blur and mackle problems in the overlap area using tradition linear weight function. Thus we can gain pleasure visual effect using these two methods in image mosaics with notable illumination difference.Illumination is an important factor for scene generation, we proposes a pure image based rendering method to synthesis image under arbitrary illumination condition. In this method, Singular Value Decomposition (SVD) was used analytically to derive environment light part and properties of scenes from sampled images, which were taken from the same view point under different point illumination condition. For images rendering under novel lighting conditions, direct illumination effect is calculated using lambertain model; while indirect illumination one was added by environment light part derived before. So feature of this method is that it supplies a good algorithm to synthesis image in arbitrary illumination environment.Aimed at the needs of high image resolution and zooming browse in virtual environment, this paper research algorithm of image resolution enhancement. In the case of one frame, we presented a novel image zooming algorithm based on maximum a posteriori (MAP) probability theory and maximum entropy principle, which anisotropy interpolate pixels using gradient information self-adaptively. In the case of multi-frames, we presented a super-resolution reconstruction algorithm based on generalized MAP theory and maximum entropy principle. By introduce bifilter into traditional MAP algorithm, we anisotropy interpolate pixels using gradient information and geometry distance self-adaptively. Experiment shows that our algorithm holds higher signal/noise ratio and better vision quality.Aimed at the needs of virtualized relationship in virtual environment, we introduce non-photorealistic rendering into virtual reality and presented a novel algorithm for virtualized relationship generation in image based virtual environment based on line integral convolution. First, structure tensor is calculated to gain vector field for characterizing local structure of image. And then, total variation partial difference equation is imposed for smoothing vector field. Based on this we use line integral convolution for guide the stroke to generate virtual special effect image with fluctuant sense. Last, mergering with original real image, cylindrical panorama with virtualized relationship controlled by time can be generated, which accomplished the character of surpass the time and space in virtual reality system.Aimed at the needs of divagation in virtual environment, we define the position chain, time chain, zooming chain and virtulized chain. Also, a divagation strategy is proposed for vitural panorama space containing position relationship, time relationship, zooming relationship and virtualized relationship.更多还原