基于移动最小二乘法的视差图像拼接

PARALLAX IMAGE STITCHING BASED ON MOVING LEAST SQUARE METHOD

对于视差图像拼接,现有的工作大都采用单应性变换,这不足以得到好的拼接结果。提出一个新的视差图像拼接算法。首先检测图像的特征点并匹配,随后用随机采样一致算法RANSAC(Random Sample Consensus)和距离相似性筛选出正确的匹配点集;其次,以这些特征点结合移动最小二乘法构造一个全局仿射变换对准图像;最后,在图像的重叠区域以像素为顶点构建一个网络流,用最大流最小割算法寻找最优拼接曲线,并融合图像。由于提高了特征点匹配的正确性,对准模型的准确性明显好于以前的工作,图像拼接结果平滑真实,无扭曲、鬼影等现象。

Parallax image st itching usual ly ut il izes the feature-based transformation. Existing works focused on the image transformation model, especially a homography, but this is inadequate to obtain better alignment. This paper presents a new stitching method. Firstly, detect and match features in images, then RANSAC and distance similarity are used to investigate the accuracy of matching features. Then, based on these matching point pairs, moving least squares is used to construct a global affine transformation to align two images and get the overlapped region of them. Finally, constructing a network flow whose vertices are the pixels on the overlapped region, and max-flow/min-cut algorithm is adopted to search for a best seam, then stitch and blend images according to this seam. As our method effectively strengthens the matching reliability between SIFT points, the alignment accuracy of two images is better than previous method. Image stitching result is smooth and authentic, no distortion, ghosting or other artifact.