【作者】 杨占龙；

【导师】 郭宝龙；

【作者基本信息】 西安电子科技大学 ， 电路与系统， 2008， 博士

【摘要】 图像拼接技术是计算机视觉、图像处理和计算机图形学的研究热点,它可以用来建立大视角的高分辨率图像,在虚拟现实领域、医学图像处理领域、遥感技术领域和军事领域中均有广泛的应用。图像拼接就是将同一场景的相互有部分重叠的一系列图片拼接成大幅的、宽视角的、与原始图像接近且失真小、没有明显缝合线的高分辨率图像。图像拼接的质量主要依赖于图像的配准精度。目前,基于特征点匹配的图像配准是一种主要针对仿射变换模型和透视变换模型的图像配准技术,它不仅能够适用于重合度较低的图像之间的配准,而且还能够应用于运动场景图像以及含有遮盖物体图像之间的配准,在实际应用中得到了广泛的使用。因此,研究基于特征点的图像配准和拼接技术具有重要的理论意义和实用价值。本文主要以特征点匹配为主线,针对目前流行的Harris特征点和SIFT（Scale Invariant Feature Transform）特征点匹配技术进行了深入系统地研究。主要研究工作及贡献如下:1.研究并分析了摄像机成像几何的基本原理,重点对8参数透视变换矩阵的存在条件以及8参数透视矩阵退化为6参数仿射矩阵的条件进行了论证,由此引出了图像拼接中几种常用的变换模型―相似变换、仿射变换以及透视变换,给出了每种模型的实际应用场合。2.针对传统的特征点匹配算法对于图像旋转变化敏感的问题,提出了一种基于特征点旋转归一化的图像配准算法（IPNR）。系统研究了Harris特征点的基本原理以及几种传统的匹配方法,包括像素差的平方和（SSD）、像素的互相关信息（CC）、归一化互相关信息（NCC）。分析比较了各种匹配方法的性能以及存在的不足之处。提出了在特征点匹配过程中将邻域窗口进行旋转归一化,有效解决了特征点匹配对于图像旋转变化敏感的问题。仿真结果表明,该算法能有效地克服传统特征点匹配算法对图像旋转变化敏感的问题,将正确匹配的概率提高30%以上,效果明显优于传统的特征点匹配算法。3.针对传统的特征点匹配方法对噪声敏感的问题,提出了一种基于特征点不变矩的图像配准算法（IPIM）。图像的矩特征是一种以图像分布的各阶矩来描述灰度的统计特性的方法,它对噪声和光照的变化不太敏感,并具有旋转和尺度不变性。对图像的矩特征研究发现,传统的Hu氏矩中φ₃,φ₅两个矩对于图像的旋转变化是不稳定的,因此对Hu氏矩进行了改进,并在此基础上,利用改进后的Hu氏矩作为特征点的描述符进行匹配,解决了传统特征点匹配方法对于旋转和噪声敏感的问题。仿真结果表明该算法具有较好的旋转不变性和抗噪声性能,匹配效果优于传统的算法。4.针对IPIM方法对图像亮度变化敏感的问题,提出了一种基于特征点伪Zernike矩的图像配准算法（IPPZM）。由于伪Zernike矩比Hu矩具有更多的矩数量以及更好的抗噪声性能,本算法选用特征点邻域窗口的伪Zernike矩作为特征点描述符。数字图像的像素离散性质使得伪Zernike矩的计算产生误差,不同阶矩的计算精确度不同。因此,为保证描述符的准确性与较低的计算复杂度,需要对伪Zernike矩进行优化选择,并对矩的选取数量加以限制。仿真结果表明该算法不仅有效解决了IPIM算法对亮度变化敏感的问题,而且在旋转不变性和抗噪声性能上比IPIM算法有了进一步的提高。5.针对Harris特征点对于尺度变化敏感的问题,提出了一种改进的基于SIFT特征的图像配准算法。该算法借助于SIFT特征对于旋转和尺度的不变性以及对于噪声、视角变化和光照变化等良好的鲁棒性,解决了Harris特征检测对于尺度变化敏感的问题,使得较大尺度变化下的图像配准成为可能。同时,对SIFT特征点的提取方法进行了改进,预先去除了部分较不稳定的特征点,提高了匹配的速度和正确匹配的概率。实验结果证明该算法对于旋转、尺度变化均具有不变性、对于噪声以及图像亮度变化具有较好的鲁棒性,且匹配速度比改进前提高了近1倍。6.针对图像拼接中累计误差对合成图像质量影响严重的问题,提出了一种基于特征点的整体优化调整方法。如果仅利用图像局部配准的结果对图像序列进行拼接,则会因为误差的累计而造成合成图像中出现重影或者图像变得模糊。该方法利用基于特征点的图像配准算法中得到的正确匹配点,对具有匹配关系所有图像的变换矩阵进行整体优化和调整,使累计误差降低。实验结果表明,该方法可以有效降低图像序列中的累计误差,提高最终合成图像的质量。更多还原

【Abstract】 Image mosaic is an important research area of computer vision, image processing and computer graphics. It can used to construct the high-resolution image with large angle of view, and has been widely used in virtual reality, medical image processing, remote sensing and military affairs. Image mosaic is the technique to stitch a series of overlapped pictures into a bigger one, which is a super-resolution picture with wide eyeshot, nearly no distortion to the originals and no visible seam line. The quality of image mosaic mainly depends on the precision of image registration. Currently, feature point based registration is a kind of image registration technique which mainly conforms to affine transformation model and perspective transformation model. It can be used not only in the less overlapped pictures registration, but also pictures with motion scenes or even with covering parts. Therefore, it has been widely used in practical applications. As a result, it is of great theoretic and practical value to do further researches in feature point based image registration and mosaic. In this dissertation, the author mainly focuses on the research of feature point matching of the Harris feature points and the SIFT feature points. The main research work in the dissertation is as follows::1. The basic geometrical principles of the camera is studied, with emphasis on the existence term of the eight-parameter perspective transform matrix and the term on how to obtain six-parameter affine matrix from the eight-parameter perspective matrix. Thereafter, several commonly used transform model in image mosaic, i.e., similar transform, affine transform and perspective transform, are deduced together with the practical application situation of each model.2. Aiming to solve the problem that traditional feature point matching algorithms are sensitive to image rotation, a new image registration scheme is presented based on rotation normalization and feature point (IPNP). The basic principles of the Harris interest point detector is studied systematically together with several Harris feature based descriptors, including the Sum of Squared Differences (SSD), Cross Correlation (CC) and Normalized Cross Correlation (NCC). The performances of the above descriptors when applied in feature point matching are evaluated. The scope of application and the deficiency for each descriptor are also analyzed. Then an image registration algorithm is proposed based on Harris interest point and image normalization. In the feature matching process, the neighboring area of each feature point is first rotation normalized. As a result, the effect of image rotation on feature matching is solved. Experimental results show that the proposed scheme can solve the rotation sensitive problem of traditional methods and improve the probability of correct matching by 30%, which is advantageous over the traditional methods.3. Aiming to solve the problem that traditional feature point matching schemes are sensitive to added noise, an image registration scheme is proposed based on feature point and invariant moment (IPIM). Image moment is a kind of method to describe the statistical features of an image. It is insensitive to the changes in noise and illumination. Besides, it is invariant to image rotation and scale change. Studies on image moment features show that ? 3 and ? 5 in the Hu moment is instable when the image is rotated. Therefore, the Hu moment is improved, and the feature descriptors are generated using the improved Hu moments, which are then used for feature matching. As a result, the problem that traditional feature matching methods are sensitive to image rotation and added noise is solved. Simulation results show that the proposed scheme performs well on rotation invariance and noise insensitiveness, and the matching effect is advantageous over traditional methods.4. The IPIM method is sensitive to image intensity change. In order to solve the problem, a new image registration method based on interest point and pseudo-Zernike moment is presented (IPPZM). As pseudo-Zernike moment has more number of independent moments and better ability to resist noise, the proposed scheme employs the pseudo-Zernike moments of the neighboring area centering at the feature point as the feature descriptor. The discrete nature of digital image results in the computation error of pseudo-Zernike moments. As a result, different moments are computed with different accuracies. In order to guarantee the accuracy of the descriptor and low computation complexity, an optimization procedure is necessary to select the pseudo-Zernike moments and apply a limit on the number of moments. Simulation results show that the proposed scheme not only solves the illumination sensitiveness problem, but also improves in regard of rotation invariance and noise resistance.5. The Harris feature point is sensitive to image scale change. To overcome this problem, an improved SIFT based image registration scheme is presented. The proposed scheme solve the above problem by using the rotation and scaling invariant property of the SIFT feature points as well as its robustness to added noise, viewpoint change and illumination change, which makes it possible to achieve successful image registration when large scale change occurs. Meantime, we have improved the SIFT feature point extraction method by reducing in advance some unstable feature points, so that the matching speed and precision of correct matching are improved. Experimental results show that the proposed scheme is invariant to rotation, scale change, and it has good robustness to noise and intensity change. Besides, the matching speed is improved by about 2 times.6. The accumulative errors in image mosaic systems have great effect on the quality of the synthesized image. Aiming to solve the problem, an overall optimization adjusting method is proposed based on feature points. If only the local image registration results are employed to align the image sequences, the shadow or fussy effect will occur in the synthesized image due to the error accumulation. The proposed scheme optimizes and adjusts the transform matrix using the correct feature point pairs obtained from the feature point based image registration, aiming to reduce the accumulative error. Experimental results show that the proposed scheme can effectively reduce the accumulative error between the image sequences and improve the quality of the synthesized image.更多还原