结合仿射变换及梯度描述符的密集匹配方法

A dense matching method combining affine transform and gradient descriptor

针对SIFT算法难以取得密集匹配结果,无法满足高精度建模需求,以及传统密集匹配算法对明暗变换影像较为敏感,无法取得稳定可靠的密集匹配点的问题,该文提出一种结合仿射变换及梯度描述符的密集匹配方法。该方法首先依靠SIFT算法获取稀疏匹配点,并利用RANSAC方法去除错误匹配点建立同名三角网,利用仿射变换对同名三角形各边中点进行加密,再对新加入的同名点进行相似性判定。与传统密集匹配算法中相似性判定依靠影像灰度层面不同,该文对影像的梯度层面信息进行统计并建立梯度描述符,通过阈值判定是否接受为同名点并不断更新同名三角网,以不再产生新的同名点为终止条件。实验选取6组不同类型影像,结果表明该方法不但可以得到稳定的密集匹配结果,而且较好地解决了传统密集匹配算法面对明暗变换影像匹配无力的问题,并对不同类型影像有着较好的适应性、鲁棒性。

Aiming at the problem that scale invariant feature transform(SIFT)algorithm couldn’t achieve dense matching results so that it couldn’t meet the requirements of high-precision modeling and traditional dense matching algorithms couldn’t get stable and reliable dense matching points because they are sensitive to the light-dark transform images.A dense matching method combining affine transformation and gradient descriptor was proposed.The method relied on SIFT algorithm to obtain the sparse matching points and took RANSAC method to remove the wrong matching points to establish namesake triangle networks.Affine transformation was used to encrypt the midpoints of each side of the namesake triangles,and then judged the similarity of the newly added points.Different from the similarity judgment in the traditional dense matching algorithms,the gradient level information of the image was counted to establish the gradient descriptor,and the threshold value was used to determine whether the homonymy points were accepted or not,and the namesake triangular networks were constantly updated,with no new homonymy points creating as the termination condition.6 groups of different types of images were selected in the experiment.The results showed that this method could not only obtain stable and dense matching results,but also solve the problem of the traditional algorithm’s inability to match the light-dark transformation images,and had good adaptability and robustness to different types of images.