中文摘要：

针对SIFT （scale invariant feature transform）算子在大幅复杂图像中提取的过多不稳定特征点及在只有少量重合区域下图像配准过程中出现的过多误匹配,导致图像配准精度下降;提出一种改进的SIFT算法,在对目标图像提取SIFT特征后,利用双向BBF（Best-Bin-First）匹配算法对提取的特征点进行匹配,采用SIFT描述子的尺度以及梯度方向信息建立最小邻域匹配剔除误匹配点,通过随机抽取一致性算法（RANSAC）进一步筛选匹配点,并利用最小二乘法结合多项式近似拟合出变换模型,利用局部均方根有效值（RMS）评价映射矩阵与实际图像的误差,找出并删除引起误差的误匹配点,迭代至配准图像符合评价标准后,计算出精确变换模型.实验结果表明,该算法提高了大幅复杂图像在少量重合区域时的配准精度.

英文摘要：

For the SIFT operator in big complex images, extracted many unstable feature points and many false matches when the images were only small overlap regions, which leads to a decrease accuracy in image registration. This paper proposes an improved SIFT algorithm. After extracting SIFT feature from the target image, bidirectional BBF （Best--Bin--First） algorithm is used to match the extracted features, then using the Neighbor Feature Matching eliminate false matches based on scale and gradient direction information from SIFT descriptor. Using Random Sample Consensus algorithm （RANSAC） further to filter the matching points and calculating transformation model combined with the least square method and polynomial approximation fitting. Finally, the local root mean square value （RMS） can evaluate the mapping error matrix to the actual image. If the error is bigger, the system will give the feedback and find the error matching, and then iterative to in line with the evaluation. The experimental results show the algorithm has improved the accuracy of registration, especially for big images with smaller overlap regions.