«上一篇
 文章快速检索 高级检索

 应用科技  2019, Vol. 46 Issue (2): 94-97, 103  DOI: 10.11991/yykj.201806011 0

引用本文

YANG Fujia, ZHENG Liying. New feature extraction matching algorithm based on SIFT[J]. Applied Science and Technology, 2019, 46(2), 94-97, 103. DOI: 10.11991/yykj.201806011.

文章历史

New feature extraction matching algorithm based on SIFT
YANG Fujia, ZHENG Liying
College of Computer Science and Technology, Harbin Engineering University, Harbin 150001, China
Abstract: The traditional image matching algorithm has the problems of unstable feature points and slow matching time. Therefore, this paper proposes an improved image matching algorithm based on scale invariant feature transform （SIFT）. Firstly, the Gaussian multi-scale space was constructed for the traditional Harris corners, so that the corners have multi-scale invariant characteristics. Then, the Canny edge extraction algorithm was used to modify Harris corners to increase the number of stable feature points. Finally, SIFT feature descriptors were constructed to calculate the Euclidean distances of the corresponding feature point descriptors in multiple images and complete the matching of feature point pairs. Experimental results show that compared with the traditional SIFT algorithm and SURF algorithm, the prop osed method can effectively improve the matching accuracy of feature points and reduce the image matching time.
Keywords: image matching    SIFT algorithm    scale space    Harris algorithm    Canny edge extraction operator    feature descriptor    Euclidean distance    matching accuracy

1 SIFT算法

SIFT算法被认为是图像匹配效果最好的方法之一，它对物体的尺度变化、刚体变换、光照强度和遮挡都具有较好的稳定性。SIFT算法总共可分为4个阶段：尺度空间构建、特征点选择、方向确定和特征点描述。尺度空间构建的基本思想是在输入的图像模型中，通过高斯模糊函数连续地对尺度进行参数变换，最终得到多尺度空间序列。图像中某一尺度的空间函数Lxyσ）由可变参数的高斯函数Gxyσ）和原输入图像Ixy）卷积得出

 $L(x, y, \sigma ) = G(x, y, \sigma ) * I(x, y)$

 $G(x, y, \sigma ) = \frac{1}{{2{\text π} {\sigma ^2}}}{e^{ - ({x^2} + {y^2})/2{\sigma ^2}}}$

 \begin{aligned}f_{\rm DOG}(x, y, \sigma ) = & [G(x, y, n\sigma ) - G(x, y, \sigma )] * I(x, y)= \\ & L(x, y, n\sigma ) - L(x, y, \sigma ) \end{aligned}

 $F(x, y) \!\!=\!\! \sqrt {{{{\text{(}}L(x \!\!+\!\! 1, y) - L(x \!\!-\!\! 1, y){\text{)}}}^{\text{2}}} \!\!+\!\! {{(L(x, y \!\!+\!\! 1) - L(x, y \!\!-\!\! 1))}^2}}$ (1)
 $H(x, y) = {\tan ^{ - 1}}(\frac{{L(x + 1, y) - L(x - 1, y)}}{{L(x, y + 1) - L(x, y - 1)}})$ (2)

2 改进的方法

2.1 多尺度Harris角点提取

 $M(x, y, \sigma ) = \sum\limits_{x, y} {G(x, y, \sigma )} \left[ {\begin{array}{*{20}{c}} {I_x^2}&{{I_x}{I_y}} \\ {{I_x}{I_y}}&{I_y^2} \end{array}} \right]$

 $Q(x, y) = \det M - k \times {\rm trace}\;{M^2}$

2.2 Canny边缘提取

 $E(E) = \sqrt {I_x^2 + I_y^2}$
 $D(\theta ) = {\tan ^{ - 1}}(\frac{{{I_x}}}{{{I_y}}})$

2.3 特征点描述和匹配

3 实验结果与分析