基于尺度空间理论的图像特征提取技术研究
发布时间：2018-07-29 08:29
【摘要】：随着图像特征提取技术在计算机视觉领域的应用越来越重要,对能够提取鲁棒性更好,更能反映图像内容属性的图像特征的要求越来越高。与全局特征相比,图像局部特征对图像的几何与光学变换(如旋转、尺度、仿射、光照变换等)更鲁棒,除此之外,局部特征还具有很高的重复性,并且不容易受到目标遮挡的影响,所以对图像局部特征提取技术的研究越来越受到重视。在图像匹配应用中,为了提高图像特征的综合性能(尺度、仿射不变性、实时性),本论文对图像局部特征提取技术做了详细的研究和深入分析。本文首先深入研究了尺度空间,在此基础上详细研究了基于尺度空间的图像特征检测算法的基本原理,包括:基于高斯差分算子(Difference of Gaussian,Do G)的SIFT检测器,基于高斯尺度空间的Harris-Laplace尺度不变检测器以及能够应用于存在仿射变换情况的扩展算法Harris-Affine和Hessian-Affine仿射不变特征检测器。尺度空间在特征检测中的应用保证了检测到的特征具有尺度不变性以及更好的稳定性。并且通过仿真实验对比了各种检测算法的性能,基于实验结果我们选择鲁棒性最优的Hessian-Affine检测器作为后续特征描述的预处理算法。其次,本文从图像匹配的角度,深入研究了图像特征描述算法,包括MROGH,FRDOH和LIOP算法,通过对算法原理的研究,得出通过引入多支撑域思想或者双梯度直方图思想能够提高特征描述符的鲁棒性和可区别力的结论。但是引入多支撑域构造描述符会增加构造过程所消耗的计算时间,降低实时性。最后我们提出了一种基于双梯度方向直方图的特征描述新方法(DGOH),并且通过实验对DGOH算法与其他算法进行比较可以得出,DGOH算法的鲁棒性优于FRDOH,并且与其他算法相比DGOH具有最优的实时性能。
[Abstract]:With the application of image feature extraction technology in the field of computer vision, the requirements of image features which can extract better robustness and more reflect the image content attributes are becoming more and more important. Compared with global features, image local features are more robust to geometric and optical transformations (such as rotation, scale, affine, illumination transformation, etc.). In addition, local features are highly reproducible. And it is not easy to be affected by object occlusion, so more and more attention has been paid to local feature extraction. In the application of image matching, in order to improve the comprehensive performance of image features (scale, affine invariance, real-time), this paper makes a detailed study and in-depth analysis of image local feature extraction technology. In this paper, the basic principle of image feature detection algorithm based on scale space is studied in detail, including: SIFT detector based on Gao Si differential operator (Difference of Gaussian do G). Harris-Laplace scale invariant detector based on Gao Si scale space and extended algorithm Harris-Affine and Hessian-Affine affine invariant feature detector which can be applied to the existence of affine transformation. The application of scale space in feature detection ensures that the detected features are scale-invariant and more stable. The performance of various detection algorithms is compared by simulation experiments. Based on the experimental results, we choose the robust Hessian-Affine detector as the preprocessing algorithm for the subsequent feature description. Secondly, from the point of view of image matching, this paper deeply studies image feature description algorithm, including MROGH FRDOH and LIOP algorithm. It is concluded that the robustness and distinguishing force of feature descriptors can be improved by introducing the idea of multi-support domain or double-gradient histogram. But the introduction of multi-support domain construction descriptor will increase the computation time and reduce the real-time performance. Finally, we propose a new feature description method based on double gradient histogram (DGOH),). By comparing the DGOH algorithm with other algorithms, we can conclude that the robustness of DGOH algorithm is better than that of other algorithms, and compared with other algorithms. DGOH has optimal real-time performance.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2152094
[Abstract]:With the application of image feature extraction technology in the field of computer vision, the requirements of image features which can extract better robustness and more reflect the image content attributes are becoming more and more important. Compared with global features, image local features are more robust to geometric and optical transformations (such as rotation, scale, affine, illumination transformation, etc.). In addition, local features are highly reproducible. And it is not easy to be affected by object occlusion, so more and more attention has been paid to local feature extraction. In the application of image matching, in order to improve the comprehensive performance of image features (scale, affine invariance, real-time), this paper makes a detailed study and in-depth analysis of image local feature extraction technology. In this paper, the basic principle of image feature detection algorithm based on scale space is studied in detail, including: SIFT detector based on Gao Si differential operator (Difference of Gaussian do G). Harris-Laplace scale invariant detector based on Gao Si scale space and extended algorithm Harris-Affine and Hessian-Affine affine invariant feature detector which can be applied to the existence of affine transformation. The application of scale space in feature detection ensures that the detected features are scale-invariant and more stable. The performance of various detection algorithms is compared by simulation experiments. Based on the experimental results, we choose the robust Hessian-Affine detector as the preprocessing algorithm for the subsequent feature description. Secondly, from the point of view of image matching, this paper deeply studies image feature description algorithm, including MROGH FRDOH and LIOP algorithm. It is concluded that the robustness and distinguishing force of feature descriptors can be improved by introducing the idea of multi-support domain or double-gradient histogram. But the introduction of multi-support domain construction descriptor will increase the computation time and reduce the real-time performance. Finally, we propose a new feature description method based on double gradient histogram (DGOH),). By comparing the DGOH algorithm with other algorithms, we can conclude that the robustness of DGOH algorithm is better than that of other algorithms, and compared with other algorithms. DGOH has optimal real-time performance.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前1条
1 刘美玲;高虎明;;基于Harris-Laplace特征点检测抗几何攻击的数字水印算法[J];科学技术与工程;2014年10期
相关博士学位论文 前3条
1 刘景能;图像局部不变特征提取技术及其应用研究[D];上海交通大学;2012年
2 高彦彦;基于特征的图像检索技术研究[D];北京邮电大学;2012年
3 胡俊华;图象局部不变特征及其应用研究[D];中国科学技术大学;2009年
相关硕士学位论文 前2条
1 刘美玲;基于Harris-Laplace特征点检测抗几何攻击的数字水印算法研究[D];天津财经大学;2014年
2 侯毅;图像局部不变特征提取及应用研究[D];国防科学技术大学;2012年
，本文编号：2152094
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2152094.html
