Abstract: In order to improve the performance of the image features correspondence that solely depends on similarity, a correspondence algorithm is proposed, which used the geometric constraint efficiently by the voting method. First of all, a fast coarse matching by the descriptor of the features was performed to get the candidate correspondence, then a robust geometric constraint was used to compute a new similarity of the candidate correspondence by the voting method. Finally, a self-adaptive method for distinguishing the belief of the candidate correspondence was proposed to get the correspondence of image features. Comprehensive experiments demonstrate that our method is effective and robust for a variety of image transforms. Better result can be got by the correspondence algorithm based on voting of geometric constraint than that based on similarity simply.

Key words: image feature, features correspondence, geometric constraint, voting, adaptive threshold

摘要： 为了改进单纯依靠相似度的图像特征匹配效果，提出了一种特征匹配方法，采用投票的方法在特征匹配的过程中引入几何约束。首先根据特征的描述向量进行初步的快速匹配，得到候选特征对以后通过投票的手段在特征之间相似度矩阵中引入一种鲁棒的几何约束，并通过自适应阈值过滤的方法获得图像特征匹配。在实验中验证了该方法对图像特征匹配正确与否具有较强的区分能力，在与已有技术接近的计算时间内得到了更高的匹配准确度。基于几何约束投票的图像特征匹配比单纯依靠相似度具有更好的匹配效果。

关键词: 图像特征, 特征匹配, 几何约束, 投票, 自适应阈值