基于稠密局部自相似特征流的图像配准算法

An image registration algorithm based on dense local self-similarity feature flow

针对基于稠密SIFT流图像配准算法执行效率和配准准确率有待提高的不足,提出了一种基于稠密局部自相似(LSS)描述符构建的稠密改进的LSS(ILSS)特征流的图像配准算法。算法通过颜色空间转换分离出彩色图像中的颜色和亮度信息,只在亮度通道上提取稠密LSS特征以大幅度提高图像特征提取阶段执行效率。随后以保持特征流场光滑性为约束条件,采用金字塔多分辨率迭代法提高LSS特征流场估计阶段的执行效率。多分辨率迭代法的基本思想是先在图像粗粒度网格上快速估算出初步的特征流场,然后再逐步求精获得最终精确的特征流场。大量实验表明,与稠密SIFT流相比,基于稠密ILSS特征流的图像配准算法在图像内容发生较大变化时具有更好的鲁棒性,同时具有更高的执行效率和图像配准确率。

Aiming at pro.rooting execution efficiency and registration accuracy of dense scale invariant feature transform （SIFT） flow based image registration algorithm, in this paper, a new improved dense local self-similarity （ILSS） flow based registration algorithm is proposed. By separating color and luminance information of color images through the color space conversion, the improved image registration algo- rithm only extracts dense self-similar feature on the channel of luminance to improve efficiency of image feature extraction stage substantially. With constraints of maintaining the characteristics of the flow field smoothness, we then employ pyramid multi-resolution iterative method to improve the computational el- ficiency of the estimation of the self-similar flow field stage. The basic idea of the multi-resolution itera- tire method is to roughly estimate the dense ILSS flow at a coarse level of image grid, then gradually propagate and refine the dense flow from coarse to fine. A large number of experimental results show that compared with dense SIFT flow based image registration algorithm,the dense ILSS flow based one allows robust matching across different scene appearances and has higher computational efficiency and registration accuracy. The success on these experiments suggests that the image registration using dense ILSS flow can be a useful tool for various applications in computer vision and computer graphics.