《图像拼接系统算法研究及实现.doc》由会员分享，可在线阅读，更多相关《图像拼接系统算法研究及实现.doc（108页珍藏版）》请在金锄头文库上搜索。
1、四川大学硕 士 学 位 论 文题 目 图像拼接系统算法研究及实现 作 者 华骅 完成日期 2007年5月11日 培 养 单 位 四川大学 指 导 教 师 罗代升 专 业 电路与系统专业 研 究 方 向 智能系统与设计 授予学位日期 2007 年 6 月 日 - 105 -图像拼接系统算法研究及实现电路与系统专业研究生 华骅 指导老师 罗代升 教授摘要：随着数字成像设备的普及，人们对大视角图像有了越来越迫切的需求。获取这种图像有两种方法，第一是用专业设备拍摄，第二是多幅图像拼接，前者几何失真大难以校正，细节不足；后者几何失真小，细节充分，但难以拼接。本文按照实现一个图像拼接系统所需步骤，首先针对
2、传统基于像素的匹配方法和使用Harris角点法的不足，使用SIFT算法实现了旋转不变、放缩不变，而且能够适应一定光照和视点变化的特征提取。为获得各幅图像之间重叠部分的关系，需对提出的特征进行匹配，针对传统最近邻查找方法速度缓慢的缺点，提出了采用位置敏感散列算法，提高了匹配速度，并详细讨论了它的原理和实现细节。还将随机选取一致性算法用于错误匹配点对的筛除，解决了匹配结果会有很多错匹配的问题，可以获得完全正确的匹配点对。通过对成像几何的分析，说明了基于变换矩阵的拼接方法，并指出它的不足和解决办法。为从根本上避免变换矩阵法的不足，在深入研究基于拍摄模型的方法后，建立了球体模型的拼接流程，并使用一种普
3、遍适应的扫描模型，在统一的优化框架下解决了扫描图像拼接问题。还讨论了多种图像插值方法，并对拍摄模型法进行了拓展，使其适应性能更佳，鲁棒性更好。研究了多种拼接缝消除技术，揭示了在图像间有较大光照或色彩差异的情况下仅采用拼接缝消除技术所带来的不足，针对有较大光照或色彩差异的情况，采用了一种基于直方图的光照和色彩平衡方法，进行统一调整，然后再使用接缝消除技术能得到效果良好的拼接图像。讨论了全自动图像拼接系统所需要的技术细节和实现方法，实验表明本系统具有很好的鲁棒性和自动拼接能力。关键词：图像拼接 SIFT算法 位置敏感散列算法 鲁棒估计 球体模型 图像融合Study on Image Mosaic
4、SystemCircuit and SystemGraduate Student: HUA Hua Advisor: Luo DaishengABSTRACT With the popularity of digital imaging equipment, image with wide field of view are need more and more urgent. There are two methods to get such images, the first is that image with wide field are captured with professio
5、nal equipment, and the second is get the images with field with the technology of image mosaic. The geometric distortion of the former is large, hard to correlation and has insufficient details, and that of the latter is small and has full details. However the latter is difficult to mosaic.In this p
6、aper, according to the realization step for a Image Stitching System, first we point out the inadequate of traditional pixel-based matching methods and the Harris corner method, use SIFT algorithm to extract interest points which can achieve rotate invariant and scale invariant, and able to adapt to
7、 some changes in light and viewpoint.To obtain the image overlap between parts of the relationship, it is necessary to match the interest point. The traditional method of nearest neighbor search has very slow speed, to overcome the shortcoming, Locality Sensitive Hashing algorithm is used, and repre
8、sent detailed discussion of its principles and implementation details. In addition, Random Sample Consensus method is being used to remove mismatching interest points, in the result we get the complete correct matching point pairs.Based on the analysis of imaging geometry, we represent homography ma
9、trix stitching, and point out its shortcomings and solutions. In order to avoid homography matrix method shortcomings fundamentally, after in-depth study on shoot model method, a sphere stitching workflow is being established. And presents a general model of scanning image stitching which can adapt
10、to many scan-like situatons. Achieve a unified framework to resolve the problem of scanning images. Also discussed a wide range of image interpolation method, and expand shooting model method to make better adapability and robustness.We study some seamlines eliminating technicals, and reveal the ins
11、ufficiency of the way which only uses seamlines eliminating technical to the image in a large light or color difference. In order to deal with large differences of illumination and color situation, a method based on light and color histogram balance is proposed. Use that method to align all images i
12、nto same illumination or color situation, and then use seamlines eliminating technical will achieve excellent result.Discussed the technical details and implement methods of automatic image stitching system, the experiments show that the system is robust and has automatic stitching capability.Keywor
13、ds image mosaic, SIFT algorithm, Locality Sensitive Hashing algorithm, robust estimate, sphere model, image blending目 录1.绪论11.1.课题研究的意义与背景11.2.图像拼接技术的现状21.3.本课题的主要工作41.4.论文的组织结构52.图像特征提取62.1.尺度空间理论62.2.SIFT算法82.2.1.尺度空间极值提取82.2.2.亚像素精度122.2.3.候选特征点筛选142.2.4.候选特征方向赋值162.2.5.提取特征点描述182.2.6.算法流程212.3.本
14、章小节213.特征点匹配233.1.最近邻查找问题233.1.1.k-d树算法253.1.2.位置敏感散列算法273.2.特征匹配筛选383.2.1.粗筛选383.2.2.鲁棒估计393.2.3.RANSAC算法403.2.4.去除错误匹配点对413.2.5.试验结果443.3.本章小结454.拼接图像的生成464.1.成像几何464.1.1.成像模型464.1.2.照相机运动494.2.基于变换矩阵的拼接514.2.1.求解变换矩阵参数524.2.2.捆绑调整534.3.基于拍摄模型的拼接544.3.1.像素焦距554.3.2.球体模型564.3.3.参数优化求解594.3.4.扫描模型64
15、4.4.拓展和细节664.4.1.模型拓展664.4.2.图像插值684.5.本章小节725.图像融合735.1.平均值法735.2.加权平均值法745.3.最佳拼接缝法755.4.多分辨率法775.4.1.高斯金字塔785.4.2.拉普拉斯金字塔795.4.3.边界处理805.4.4.算法流程815.4.5.实验结果815.5.光照和色彩平衡855.6.本章小结906.总结和展望916.1.总结916.2.展望92参考文献93科研成果98声明99学位论文版权使用授权书99致谢100图像拼接系统算法研究及实现1. 绪论1.1. 课题研究的意义与背景数字成像设备（数码相机、数码摄像机等）的普及，使得数码图像得到了越来越广泛的应用。在实际的科学研究和工程项目中，经常会用到超过人眼视角的超宽视角图像。由于距离的限制，普通数码相机的视角往往不能
