基于稀疏转稠密机制的仿射不变特征匹配算法

Affine Invariant Feature Matching Algorithm on Sparse-to-Dense Mechanism

摘要:

本文提出一种基于稀疏、稠密特征转换的仿射不变特征匹配算法，其中稀疏特征包括坐标，尺度，仿射模拟参数等，稠密特征指基于图像局部区域内光学属性的局部描述符。本文算法在Affine-SIFT算法基础之上，针对在特征提取阶段仅使用稀疏特征提取的缺陷做出了改进。由于稠密信息只有在稀疏参数满一定足检测条件时才能提取到特征，导致本可以匹配到的特征（包括稀疏、稠密参数）无法提取，本文将通过使用稀疏特征构造新的模拟图像，通过将稀疏特征重新稠密化，并在模拟图像基础上进一步提取稀疏特征，同时可检测到原始图像中检测不到的可匹配特征，最终达到增大特征建立匹配的概率，提升正确匹配数量的目标。经实验验证，本文提出的稀密特征转换算法相比于ASIFT算法能大量增加特征匹配的数量。除针对ASIFT方法提供扩展外，该方法也可用于扩展具有充分稀疏特征参数的其它特征提取和匹配方法，并适用于目标识别、目标分类和三维重建等问题。

Abstract:

In this paper, an affine invariant feature matching algorithm is based on sparse features into dense features is proposed, in which sparse features include coordinates, scale, affine parameters etc., dense features include information of Gaussian kernel, area descriptor. Based on the Affine-SIFT algorithm, this algorithm improves the shortcomings of sparse feature extraction in the feature extraction phase. Because the dense information can only extract the feature when the sparse parameter is full of enough detection conditions, it can not match the characteristic (including sparse and dense parameters) that can be matched, in this paper, we will reconstruct the sparse features by using the sparse features to construct the new simulation images, and further extract the sparse features on the basis of the simulated images, and can detect the matching features that can not be detected in the original image, Feature set the probability of matching, to improve the correct match the number of goals. Compared with the ASIFT, the algorithm, in this paper, which can significantly increase the number of correct feature matching points is proved by experiments. In addition to the extending method of the ASIFT. The proposed method can also be used to extend other feature extraction and matching methods with sufficiently parameters of sparse feature, and apply to precisely target recognition, target classification and 3D reconstruction etc.