SIFT与Harris算法在图像特征点提取中的适用性研究

发布时间：2018-08-03 15:04

【摘要】：近年来,随着数字图像处理技术的迅猛发展,图像特征匹配技术已在众多领域之中得到了大量的应用。至于图像特征匹配的首要任务就是从待匹配的图像中获得较稳定的图像特征点,接着通过构建待匹配图像与初始图像特征点集的关系来进行配准。而对于特征点的提取涉及到选择恰当的点特征提取方法,考虑到点特征的提取方法对图像配准的速度和准确度有一定的影响,因而选择合适的点特征提取的方法有一定的实用价值与理论意义。本文围绕图像特征点的提取,针对SIFT算法中降噪阈值的选值,Harris角点检测算法与SIFT算法在不同种类图像特征点提取上的适用性等进行了分析研究。具体的研究内容如下:(1)为避免常规SIFT算法提取特征点时选取降噪阈值存在的盲目性,以及不恰当的降噪阈值使SIFT算法提取到的特征点效果不佳的问题,研究了在SIFT算法提取特征点时对于降噪阈值来说是否存在相对比较稳定的值,通过选取此降噪阈值可使SIFT算法提取图像特征点的效果最佳。基于多组相同限制条件的实验图像(包括原始图像、旋转图像以及加“噪声”图像),选取不同的降噪阈值分别进行大量的实验。实验结果表明当降噪阈值近似选取在30附近,此时可以使SIFT算法获得相对比较稳定的特征点数目。(2)对SIFT算法和Harris角点检测算法提取特征点的原理分别进行了分析,至于SIFT算法运算过程较繁琐,而且其构建高斯差分尺度空间的过程比较复杂,考虑到Harris角点检测算法运算过程比较简便。为了直观的对两种算法进行比较,在这里采用了文中所提的3方面(特征点有效性、计算时效性以及特征点相似不变性),并在同一实验对象的条件下对两者进行了定量分析研究。(3)为了探究出SIFT算法和Harris角点检测算法在不同种类图像特征点提取上的适用性。本文针对两类图像:折线特征主导的图像与光滑曲线特征主导的图像,并提出了3个方面:特征点有效性、计算时效性以及特征点相似不变性,又采用了图像匹配对数比率、样本均值和标准差3个相关指标对两种算法进行了定量研究。最终实验结果表明,对于折线特征主导的图像:在特征点有效性、计算时效性以及特征点相似不变性上,Harris角点检测算法的适用性比SIFT算法更好一些;然而对于光滑曲线特征主导的图像,由于无法检测到一定数量的特征点,这也就导致Harris角点检测算法的适用性不如SIFT算法。最后,对本文的相关工作进行了总结,同时结合本文存在的不足之处,对今后的研究内容进行了展望。
[Abstract]:In recent years, with the rapid development of digital image processing technology, image feature matching technology has been widely used in many fields. The primary task of image feature matching is to obtain more stable image feature points from the image to be matched. Then the relationship between the image to be matched and the initial image feature point set is constructed to be registered. However, the extraction of feature points involves the selection of appropriate point feature extraction methods, and the method of point feature extraction has a certain effect on the speed and accuracy of image registration. Therefore, it has certain practical value and theoretical significance to select the appropriate method of point feature extraction. This paper focuses on the extraction of image feature points and analyzes the applicability of SIFT algorithm and SIFT algorithm in extracting different kinds of image feature points. The specific research contents are as follows: (1) in order to avoid the blindness of selecting the denoising threshold when the conventional SIFT algorithm extracts the feature points, and the problem that the improper denoising threshold makes the feature points extracted by the SIFT algorithm ineffective, This paper studies whether there is a relatively stable value for denoising threshold when extracting feature points in SIFT algorithm. By selecting this denoising threshold, the SIFT algorithm can get the best result of extracting image feature points. Based on the experimental images (including original images, rotating images and adding "noise" images) with the same restricted conditions, a large number of experiments were carried out with different noise reduction thresholds. The experimental results show that the SIFT algorithm can obtain relatively stable number of feature points when the denoising threshold is approximately 30. (2) the principle of extracting feature points from SIFT algorithm and Harris corner detection algorithm is analyzed respectively. As for the SIFT algorithm, the operation process is complicated, and the process of constructing Gao Si differential scale space is more complex. Considering the Harris corner detection algorithm, the operation process is relatively simple. In order to compare the two algorithms intuitively, we use the three aspects mentioned in this paper (feature point validity). (3) in order to find out the applicability of SIFT algorithm and Harris corner detection algorithm in extracting feature points of different kinds of images. In this paper, for two kinds of images: the image dominated by the broken line feature and the image dominated by the feature of smooth curve, three aspects are proposed: the validity of the feature point, the calculation of the timeliness and the similarity invariance of the feature point, and the image matching logarithmic ratio is adopted. The sample mean and standard deviation are used to quantitatively study the two algorithms. Finally, the experimental results show that the SIFT corner detection algorithm is better than SIFT algorithm in terms of feature point validity, computation time efficiency and feature point similarity invariance for images dominated by broken line features. However, for images dominated by smooth curve features, because a certain number of feature points can not be detected, the applicability of Harris corner detection algorithm is not as good as that of SIFT algorithm. Finally, this paper summarizes the related work, combined with the shortcomings of this paper, the future research content is prospected.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】