目的 针对不同视点下具有视差的待拼接图像中特征点筛选存在漏检率高和配准精度低的问题，提出了一种基于特征点平面相似性聚类的图像拼接算法。方法 首先根据相同平面特征点符合同一变换的特点计算特征点间的相似性度量，利用凝聚层次聚类把特征点划分为不同平面，筛选误匹配点。其次将图像划分为相等大小的网格，利用特征点与网格平面信息计算每个特征点的权重，通过带权重线性变换计算网格的局部单应变换矩阵。最后利用多频率融合方法融合配准图像。结果 在20个不同场景图像数据上特征点筛选比较实验， RANSAC算法的平均误筛选个数为30，平均误匹配个数为8，而本文方法的平均误筛选个数为3，平均误匹配个数为2。在20个不同场景图像拼接比较实验，与AutoStitch、APAP、AANAP等3种算法进行了图像拼接实验比较，本文算法相比性能第2的算法，PSNR平均提高了8.7%,SSIM平均提高了9.6%。 结论 该方法处理后的图像保留了更多的特征点，提高了配准精度，具有更好的拼接效果。
High-precision parallax image stitching method using feature points clustering
Xie Cong-Hua,Zhang Bing,Gao Yun-Mei(School of Computer Science and Engineering,Changshu Institute of Technology,Suzhou;Library,Changshu Institute of Technology,Suzhou)
Objective Image stitching is a technology for solving the field of view (FOV) limitation of images by stitching multiple overlapping images to generate a wide-FOV image. Parallax image stitching remains a challenging problem. Although many methods and an abundance of commercial tools are immensely useful in helping users organize and appreciate photo collections, many tools fail to give convincing results when given parallax images. Parallax image stitching methods with local homography transforms for partitioning cells are the most popular. However, many of those methods have high rate of wrongly matching of feature points and low accuracy of aligning feature points in different viewpoint images. We proposed a novel stitching method using hierarchical agglomerative clustering of the feature points with their plane similarity information to improve the precision rate of matching feature point. Method Firstly, we proposed a feature point shift algorithm using the clustering results of the feature points with the planar information. The SIFT (Scale-Invariant Feature Transform) feature points of all images with different viewpoint are extracted, and the k nearest-neighbours for each feature point are found using the KD-tree algorithm. K minimum-sample sets are constructed, in which each set includes 4 non-collinear feature points, to compute the homography matrix and residual matrix. Secondly, the planar information similarities of all feature points are computed according to the residual matrix, and all feature points are divided into different clusters by the hierarchical agglomerative clustering algorithm. The feature points in each cluster have a common plane and have the same homography transformation. If the mean of residual matrix in one cluster is larger than a thresholding, then the feature points in this cluster are labelled as the wrongly matching feature points. Secondly, we proposed an image stitching algorithm that it partitions image into cells with blend weights for multi-plane images. All images are partitioned into equal-sized cells, and the local homography transformation of each cell is computed by the linear transformation with blend weights. The weight of each feature point is computed using the plane information of the feature points. If one feature point and its cluster cent point have the same planar label, then the weight is equal to 1, else the weight is equal to their Gaussian kernel radial distance. Finally, the aligned images are rendered as panorama using a multi-band blending method. Result We compared our feature point detection algorithm with the RANSAC algorithm on traditional building and pavilion images. The RANSAC algorithm found 427 and 541 matched feature points respectively, while our algorithm found 435 and 589 matched feature points respectively. For the traditional building images, the RANSAC algorithm has 6 pairs of wrongly matched point, while our algorithm has only 1 pair of wrongly matched point. For the pavilion images, the RANSAC algorithm has up to 20 pairs of wrongly matched point, while our algorithm has only 1 pair of wrongly matched point. On other 20 different scene images, the average number of error feature points detected by the RANSAC algorithm is 30, while which of our method is 3; and the average number of wrongly matched point pairs is 8, while which of our method is 2. We compared our image stitching method with the 3 state-of-the-art methods of AutoStitch, APAP and AANAP on traditional and modern building images. AutoStitch have obvious seam line and ghosting in the results because of the global homograph, while APAP and AANAP have better results with somewhat ghosting. On the 20 different scene images, the PSNR and SSIM indexes of our method is increased by 8.7% and 9.6% of which of the second best method, respectively. Conclusion This paper has presented a novel method for high-precision parallax image stitching using feature points clustering. The experiment results show that our method constructing the local homography transformation to shift feature points with the planar information clustering results can increase the number of matched feature points, reduce the number of wrongly shifting feature points and wrongly matching feature points, and increase the precision of feature points alignment than the state-of-the-art image stitching approaches of AutoStitch, APAP and AANAP. The experiment results show that our method image stitching algorithm partitioning image into cells and with blend weights for multi-plane images can achieve better image stitching performance about pixel and image structure indexes than the state-of-the-art image stitching approaches of AutoStitch, APAP and AANAP.