多尺度特征点聚类的图像配准算法

Image Registration of Multi-scale Feature Points Clustering

针对在图像配准中多尺度角点检测时,图像的局部特征在一定的尺度范围内被多次检测到而产生冗余点问题,提出了一种基于多尺度特征点聚类的图像配准的方法,该算法利用聚类的思想将检测到的多尺度角点进行组合,选取每组角点响应值最大的点作为特征点,用SIFT描述子对特征点进行描述并匹配找出同名点,最后利用随机抽样一致算法求解出透视变换矩阵从而进行配准.实验结果表明,本算法减少了大量的冗余点,降低了匹配误差,提高图像配准的速度和精度,并且增强了对图像方向、亮度、尺度和噪声等变换条件下的鲁棒性.

Focused on the issue that the local feature of the image is detected repeatedly within a certain range of scales when detecting multi-scale corner in image registration and it will generate lots of redundant points,an improved image registration algorithm based on multi-scale feature points clustering was proposed. The algorithm grouped the detected multi-scale corners by using the idea of clustering and selected the corner point with the maximum response of each group as the feature point. Then the feature points were described with SIFT descriptor and matched. Finally,the matrix of perspective transformation was solved by using RANSAC algorithm for image registration. The experimental results show that the improved algorithm can decrease lots of redundant points,reduce the false matching and improve the accuracy and speed of image registration,besides,the robustness to image rotation,brightness,size and noise is enhanced.