基于奇异值分解的宽基线图像匹配算法

Wide-baseline Image Correspondence Based on Scale and Rotation Invariant Feature

图像匹配是计算机视觉中许多应用研究的基础。窄基线匹配技术虽然较为成熟,但是解决能力有限,不能处理较大的尺度、旋转、亮度以及仿射变化引起的宽基线图像序列的匹配。针对宽基线图像序列匹配的特点,在分析传统SVD匹配算法不足的基础上,引入具有尺度和旋转不变性的特征,改进邻近矩阵的度量方式,设计了一种新的基于奇异值分解的宽基线自动匹配算法。通过对比实验表明,该算法性能优于基于SIFT距离的匹配器和原SVD匹配算法,对于存在较大的尺度、旋转、亮度等宽基线变化的图像序列,能够自动获得更多的正确匹配点对和较高的准确性,鲁棒性强,甚至对视角和仿射变换也有一定的适应性。

Image correspondence is a key problem in computer vision. Although many matching technique in short-baseline have been developed, the wide-baseline correspondence problem with large scale, rotation, illumination and affine transformations is still not tackled very well. The paper proposed a new SVD matching method to achieve large number of accurate point correspondences between uncalibrated image sequences of the same scene for wide baseline. Normalized cross correlation between scale and rotation invariant features was introduced and the proximity matrix was redefined to improve the performance of robustness and reliability. Experimental results show that the proposed SVD matching method can be used for severe scene variations and provide evidence of improved performance with respect to the SIFT distance matcher and the previous SVD matching algorithm.