【作者】 葛亮；

【导师】 朱庆生；

【作者基本信息】 重庆大学 ， 计算机软件与理论， 2009， 博士

【摘要】 随着计算机性能的不断提高和计算机技术的不断发展,真实场景的三维模型越来越多地出现在各种计算机软件中,成为计算机软件表现现实世界的一种重要手段。相对于传统的三维模型构造手段,图像建模技术以其成本低廉、操作简单、具有高度真实感的建模方法等优点,逐渐成为计算机图形学和计算机视觉领域的研究热点。立体匹配是计算机视觉和图像建模研究中最基本的关键问题之一,该技术通过像点的视差来获取深度或距离,可以为三维重建、机器人导航等提供有用的信息。在实际应用中,因为变形、扭曲、遮挡等情况的影响,立体匹配是一个较难解决的问题。本文对图像建模中的立体匹配相关理论及一些方法进行了研究,提出了改进算法,并做了实验进行验证。论文的主要研究成果包括以下几点:在现实世界中,很多场景会包含有大块的纹理单一区域,增强纹理单一区域的立体匹配能力有利于提高立体匹配算法的性能。针对目前主流方法对图像中纹理单一区域匹配效果不佳的问题,提出了区域整体匹配的解决方案。该方案首先将图像分割成纹理单一区域和纹理丰富区域,然后将整个纹理单一区域作为匹配基元以得到纹理单一区域的稠密视差图。相对于点基元,区域基元包含的信息更多,且在图像中不易重复出现,因此可以有效减少因缺少纹理信息引起的误匹配。根据区域整体匹配的解决方案,提出了基本算法。该算法利用灰度共生矩阵描述图像的纹理特征,并通过灰度共生矩阵的对比度、熵和相关性特征来分割和匹配纹理单一区域。在求解纹理单一区域的视差值时,提出关键点的概念,有效解决了由于图像间纹理的细微差别使得匹配的纹理单一区域的大小形状不同所带来的问题。通过对国际标准图像上测试的实验,验证了算法的有效性。区域整体匹配的关键步骤是纹理单一区域的分割和匹配。针对纹理单一区域的特点,提出利用Laws纹理模板对图像纹理特征进行分析描述,然后进行基于直方图的分割,得到纹理单一区域。对于各种场景图像,通过分析比较各种Laws纹理模板组合,能够得到最好的分割效果。相对于灰度共生矩阵描述和分割纹理单一区域的方法,该方法能提高纹理单一区域的识别率和分割阈值选取的鲁棒性,这有助于提高基于区域整体匹配的基本算法的匹配精度和实用价值。光照强度发生变化会对基于灰度共生矩阵的纹理描述方法产生影响,这不利于纹理单一区域的分割和匹配。针对此问题,提出基于LBP/C纹理分析的立体匹配算法。该算法利用LBP/C算子不受光照变化的优点,提出通过判断局部对比度LC的值是否在0值附近,来确定像素点是否位于纹理单一区域中,并通过LBP的值来匹配纹理单一区域。相关实验验证了该算法能够有效地分割和匹配纹理单一区域。尺度变化会给物体的外观带来了较大变化,这将影响基于纹理特征的立体匹配算法的正确率。为了消除立体像对之间尺度变化所带来的影响,提出基于分形的多尺度图像区域立体匹配算法。该算法通过基于双毯的局部分形维数对像素点的局部纹理特征进行描述,进而统计图像区域内像素点的局部分形维数的分布特征,以此来描述图像区域的纹理特征。实验结果表明,相对于基于灰度共生矩阵的图像区域立体匹配算法,基于分形的多尺度图像区域立体匹配算法很好地匹配了尺度存在差异的图像区域,能够改进基于区域整体匹配的基本算法的性能。最后对论文进行了总结,并给出了论文的不足和今后研究工作的方向。更多还原

【Abstract】 With the development of computer performance and computer technology, more and more realistic 3-D models appear in all kinds of software and become an important way to represent the real world. Compared with traditional methods to construct the 3-D model, Image-based modeling (IBM) has the advantages of low cost, simple operation and realistic feeling; it turns into a hot spot of researches in computer graphic and computer vision gradually. Stereo matching, which obtains the depth or distance from the pixels’disparity and provides useful information for 3-D reconstruction, robot navigation, and so on, is one of the fundamental problems of computer vision and image-based modeling. However, stereo matching is a difficult problem under the influence of deformations, distortions and occlusions. This dissertation studies relevant theory and approaches of stereo matching and proposes some improved algorithms and experiments.The main research results of this dissertation are summarized as follows:There are many less-textured areas in the real world, so improving the ability of matching less-textured area can enhance the performance of stereo matching algorithm. To get better disparity result of the less-textured area of image, this dissertation proposes an area-based stereo matching strategy. First, the image is divided into several less-textured or well-textured areas. Then, the disparity of less-textured area is obtained through the matching based on regions. Compared with pixel, a region contains more information and appears less repeatedly, so it could reduce the probability of matching errors.On the basis of the area-based stereo matching strategy, this dissertation proposes a basic algorithm. It describes the image texture with gray level co-occurrence matrix (GLCM) and segments and matches the less-textured area by the contrast, entropy and correlation features obtained from the GLCM. This dissertation also introduces the concept of key point into the calculation of the less-textured area’s disparity. It solves the problem aroused by the difference of the size and the shape of the matched less-textured areas because of the tiny distinction in the images’texture. The experiment on the international standard image data shows that the proposed algorithm is effective.Segmentation and matching of the less-textured area are the critical steps in the area-based stereo matching strategy. According to the characters of less-textured area, this dissertation provides a novel approach for the segmentation of less-textured area. It describes the texture with Laws’masks and gets the less-textured area by histogram-based segmentation. The best result is obtained for various scene images by comparing various combinations of Laws’masks. Experimental results on the international standard image data show that the novel algorithm has better recognition rate of less-textured area and better robustness of choosing the dividing threshold than the previous method which describes and segments the less-textured area with GLCM. This new algorithm is helpful to increase the accuracy and usability of region-based stereo matching algorithm.The changing of brightness in image can influence the texture description based on GLCM, which is bad for the segmentation and matching of the less-textured area. To solve the problem, this dissertation presents a method based on LBP/C to search the less-textured area. LBP/C has resistance to the changing of brightness. So the novel method distinguish the less-textured pixel by the local contrast which is near by 0 firstly, and match the less-textured area by the distribution of LBP. Experiments show that the novel method can segment and match the less-textured area effectively.The appearance of objects will change dramatically in various scales, which decreases the accuracy of stereo matching algorithm. To reduce the influence, this dissertation proposes a fractal based multi-scale stereo matching algorithm. First, it describes the feature of the local texture with local fractal dimension based on double blanket, and then gets the feature of the image area by the distribution of local fractal dimension. The experiment results show that fractal based multi-scale stereo matching algorithm leads to better result on matching image area with various scale than GLCM based stereo matching algorithm. The novel algorithm can improve the performance of the area based stereo matching algorithm.Finally, the dissertation is concluded. Some problems as well as further work are also given.更多还原