欢迎访问《北京理工大学学报》自然科学中文版网站！

“欢迎访问北京理工大学学报自然科学中文版网站！”

白廷柱,侯喜报.基于SIFT算子的图像匹配算法研究[J].北京理工大学学报（自然科学版）,2013,33(6):622~627.
BAI Ting-zhu,HOU Xi-bao.An Improved Image Matching Algorithm Based on SIFT[J].Transactions of Beijing Institute of Technology,2013,33(6):622-627.

二维码（扫一下试试看！）

基于SIFT算子的图像匹配算法研究

An Improved Image Matching Algorithm Based on SIFT

投稿时间：2012-09-15

DOI：

中文关键词: SIFT 图像匹配余弦相似度方向一致性校正误匹配

English Keywords:scale invariant feature transform(SIFT) image matching cosine similarity consistency of orientation mismatching calibration

基金项目:

作者	单位
白廷柱	北京理工大学光电成像技术与系统教育部重点实验室, 北京 100081
侯喜报	北京理工大学光电成像技术与系统教育部重点实验室, 北京 100081

摘要点击次数: 12198

全文下载次数: 12891

中文摘要:

针对目前基于SIFT(scale invariant feature transform)的图像匹配算法在匹配相似区域较多的可见光图像时,匹配约束条件单一,没有有效剔除误匹配点,误匹配率高的问题,提出一种匹配改进算法,针对128维SIFT特征向量,采用距离匹配和余弦相似度匹配相结合的测度方法,利用特征点方向一致性进一步降低误匹配率. 实验结果表明:改进算法对图像的缩放、旋转、光照、噪声和小尺度的视角变换均有较好的匹配效果. 与原算法相比,在保证匹配点数和匹配时间的基础上,改进算法对旋转、缩放、噪声模糊和光照变换的误匹配率平均降低10%~20%,对于小尺度的视角变换,误匹配率平均降低5%.

English Summary:

For matching visible image with many similar regions, the original image matching algorithm based on SIFT (scale invariant feature transform) has the disadvantages of limited matching constraints, high false matching rate and difficulty to effectively remove mismatching points. To overcome the shortcomings above, an improved algorithm was proposed in which a combined measure of distance similarity matching with cosine similarity matching was adopted to dealing with 128-dimensional feature vectors. Further, the orientation consistency of the image feature points was employed to reduce the false matching rate. Experimental results show that the proposed algorithm has a good matching result on the conditions of image zooming, rotating, lighting, noising and small-scale perspective transformation. Compared with the original algorithm, based on the premise of ensuring enough matching points and definite matching time, the improved algorithm achieves a 10% to 20% average reduction of the false matching rate for images zooming, rotating, lighting, noising transformation and 5% for small-scale perspective transformation.

查看全文查看/发表评论下载PDF阅读器