Abstract: Since the SIFT matching algorithm and the SIFT combined with the RANSAC matching algorithm both exist the mismatching problem in varies degree, a double threshold matching algorithm based on local SIFT feature points is proposed. This paper designs the iteration criteria of the variable step size to obtain the double threshold of the SIFT, where the large threshold matching obtains a set of sparse precision matching, and the small threshold matching obtains a set of intensive matching in which mismatching may exists. Then the deformation constraint model is established based on the precise matching, which is the basis of removing the mismatching from the intensive matching. The transformation matrix is estimated by these correct matching points between the two images. To reduce the required time and increase efficiency of the algorithm, the most sharply changing region is extracted by analysing the changes of the image texture, which represents the whole image to do the matching operation. The experimental results indicate that the proposed matching algorithm has advantages of high accuracy, stability and rapidity in the situation that the affine changes of translation, rotation etc exist in the images.

Key words: Scale Invariant Feature Transform（SIFT） feature point matching, image registration, Euclidean distance, double threshold

摘要： 针对SIFT匹配算法和SIFT与RANSAC结合的匹配算法都存在不同程度误匹配的问题，提出一种基于局部SIFT特征点的双阈值匹配算法。设计变步长迭代准则获取SIFT双阈值，其中大阈值匹配获得一组稀疏的精确匹配，小阈值匹配获得一组可能存在误匹配的密集匹配。以精确匹配建立目标的形变约束模型，以此为基础从密集匹配中删除误匹配。通过这些正确的匹配点估计两幅图像之间的变换矩阵。为了降低算法所需时间，提高效率，通过分析图像的纹理变化，采用提取其变化最为剧烈的区域来代表整幅图像进行匹配运算。实验结果表明，该算法在图像存在平移、旋转等仿射变化情况下具有配准精度高，稳定和快速等特点。

关键词: 尺度不变特征变换（SIFT）特征点匹配, 图像配准, 欧氏距离, 双阈值