基于特征匹配的全景图像生成方法研究与实现
目 录
第1章绪论 ····································································· 错误！未定义书签。
1.1课题的研究背景及意义 ············································ 错误！未定义书签。 1.2基于特征匹配的全景图像研究现状 ····························· 错误！未定义书签。 1.3研究内容及论文结构 ··············································· 错误！未定义书签。 第2章基于SURF特征探测的全景图像生成方法研究 ········································· 4
2.1引言 ······························································································· 4 2.2全景图像生成算法············································································· 4 2.3SURF特征点探测算法 ········································································ 4
2.3.1尺度空间极值检测 ····································································· 4 2.3.2特征点位置确定 ········································································ 7 2.3.3特征点方向确定 ········································································ 7 2.4不变量技术特征配准算法 ··································································· 8 2.5 本章小结 ······················································································ 10 第3章基于SURF特征匹配的全景图像生成系统设计 ······································· 11
3.1引言 ····························································································· 11 3.2总体设计 ······················································································· 11 3.3 各模块设计 ··················································································· 12
3.3.1图像采集 ··············································································· 12 3.3.2图像配准 ··············································································· 13 3.3.3空白全景图像生成 ··································································· 14 3.3.4图像拼接 ··············································································· 15 3.4 界面接口设计 ················································································ 15 3.5本章小结 ······················································································· 16 第4章全景图像生成系统实现及测试 ···································· 错误！未定义书签。
4.1 引言 ···································································· 错误！未定义书签。 4.2 开发平台 ······························································ 错误！未定义书签。 4.3 关键技术实现 ························································ 错误！未定义书签。
4.3.1 SURF特征点探测 ············································ 错误！未定义书签。
III
哈尔滨工程大学学士学位论文
4.3.2对于特征变换矩阵的计算 ·························································· 18 4.4 测试 ···························································································· 18
4.4.1 SURF特征点的稳定性 ······························································ 18 4.4.2 EstimateGeometricTransform函数参数测试 ···································· 23 4.4.3程序正常运行流程及结果 ·························································· 26 4.4.4不同条件的原始图像采集对程序的影响 ········································ 30 4.5 本章小结 ······················································································ 33 结论 ······································································································· 35 参考文献 ································································································ 36 攻读学士学位期间发表的论文和取得的科研成果 ············································· 38 致谢 ······································································································· 39
IV
第1章绪论
第1章绪论
1.1 课题的研究背景及意义
图像是人类最直接的信息获取的来源，自从摄影技术诞生以来，图像技术迅速席卷全球，同时对图像的研究也进入了新的时代，其中最基本的一个问题就是图像的视野，在计算机技术出现以前，人们主要的拓展图像视野的主要方法为手动拼接和广角镜头，然而手动拼接无法应对图像之间的一些扭曲变形，相位移动等人为因素带来的痕迹，而广角镜头总会使图像产生很大的扭曲变形，不能完整真实的反应图像特征。而随着计算机技术的普及与发展，数字图像处理方式渐渐取代传统图像处理方式。
计算机的巨大计算能力使数字图像技术迅速发展壮大，让图像数字化并在此基础上进行对图像的各项操作成为了图像处理的主流。图像拼接亦得益于此。而随着计算能力的不断提高，将图像进行无缝拼接的技术越发成熟，速度也越来越快，这为自动全景无缝图像的生成奠定了基础。
全景图像是一种全方位的视图，它通过连续的变换视角获得一系列图片后，进行图像的拼接，以二维的方式展现三维，将多幅图像及图像之间的信息以单一图像的方式直观的表现出来。它最早的应用是军用遥感视图[1]。军方为了对整体进行全方位把握，需要将各个部分的遥感视图进行拼合，从而得到一幅信息量远大于单个图像的全方位视图。计算机的飞速发展使全景图像的拼接不再是军方的独有工具，借助不断发展的计算机技术，全景图像的生成门槛大大降低，一台智能手机就能轻松胜任。与此同时，正因为门槛的降低，图像拼接的应用领域大大拓宽。
图像拼接技术，特别是全景图像技术被广泛应用于航空航天[2]、军事图像分析、医学图像解析[3]、遥感图像处理[4]等多个方面，并且在计算机视觉、计算机图形学、虚拟现实等领域有着重要的课题研究。而今，除了这些专业领域方面的应用外，在我们的日常生活与工作中，它亦发挥着巨大的作用，最为浅显易见的应用就是普通移动智能客户端的图像拼接处理。
由此可见，对于全景图像的技术的研究无论是对专业领域亦或民用领域，都有着巨大的科学价值和商业价值，因此对其的深入研究有着巨大的意义。
1.2基于特征匹配的全景图像研究现状
上世纪八十年代，伴随着计算机的科技革命，数字图像处理技术已经开始了图像拼
1
哈尔滨工程大学学士学位论文
接技术的研究，经过了三十年的发展，将图像拼接技术应用到全景图像生成上已有成熟的流程，如图1.1所示，其中，最重要的一部分就是图像的配准，这一步也是影响全景图像生成的关键一步。
图1.1全景图像生成流程
在图像配准上面，目前，已有很多可以借鉴和研究的成果：
1988年，C．Harris提出了Harris角点检测算子，利用信号处理的自相关函数，在图片上找出一定量的对最小角度旋转、平移能保持不变的兴趣点[5]。但是此算法只具有一定的鲁棒性，对于图像处理预期效果而言，还有着不小的差距。因此，基于特征的匹配并没有引起更多的重视。
1994年，Blaszka通过二维高斯模糊过滤得到了一些基本的特征模型[6]，使人们更多的开始关注基于特征的图像匹配方法。
1996年，微软研究院RichardSzeliski利用LM迭代算法在其提出的八参数变换模型实现了图像的配准。这一算法亦成为图像拼接领域的经典算法[7]。同时由于其速度快，处理效果好，并且可以对抗一定程度的图像平移扭曲等多种变换，现在仍有很多人在研究该方法。
2003年，Brown和Lowe等人提出了基于尺度不变特征变换(Scale-invariant feature transform)算法[8]，并成功利用此算法实现了图像拼接[9]。该算法的提出将对图像不变特征技术的研究推向了新的高潮。对于此算法，因为其展现出了良好的鲁棒性，能够对于光照，旋转，平移，噪声，缩放甚至仿射变换进行相当程度的对抗，并且能够自动出色的完成图像的配准工作，同时它又具有相比处理同类型的其他算法而言，拥有不可比拟的速度优势，而受到了广泛的关注[10]。同时，今天的很多图像处理方法都要涉及到此算法的原理与方法，想当多的应用都基于本算法而实现的。
2006年，在SIFT算法提出不久，Bay等人通过总结前人经验加上对SIFT算法的深入研究，又提出了一种基于不变量技术的特征检测算法, Speeded Up Robust Features算法(SURF特征检测算法)[11]。该算法利用了SIFT算法的过程，通过盒子滤波与快速积分的性质，大幅度加速了特征提取的过程，同时又不损失SIFT算法的鲁棒性和抗干扰性。目前为止，此算法具有SIFT算法的大部分优势，同时在速度方面又超越了SIFT算法，因而成为图像不变量技术领域的国际领先算法。
本文所涉及的图像拼接技术正是基于SURF算法而实现的。
2