基于多特征绑定的近似重复图像检索

Near-Duplicate Image Retrieval Based on Multi-features Bundle

基于绑定特征的近似重复图像检索方法由于采用尺度不变特征变换(SIFT)特征及坐标信息,存在一定的局限性.因此,文中在SIFT特征基础上增加颜色特征形成多特征绑定,并利用兴趣点尺度方向信息,提出2种鲁棒的绑定特征兴趣点顺序计算方法:中心顺序和主方向顺序.首先,抽取绑定特征兴趣点的颜色分布特征并加入索引,检索时利用该特征的Kullback-Leibler距离排除潜在的SIFT错误匹配.然后,根据兴趣点同绑定区域中心位置的距离顺序或同主方向的夹角顺序的不一致性计算几何损失得分.在Copydays数据集上的实验表明,颜色特征的增加提升检索准确度,相比坐标顺序,绑定特征的中心顺序及主方向顺序在检索中的几何约束更好.

The bundle feature based method is effective for near-duplicate image retrieval. However, there are some limitations due to the use of SIFT descriptor and coordinate information. Therefore, in this paper color features are integrated into traditional SIFT representation to form the multi-features bundle and two kinds of robust orders between points in bundle feature are introduced-the circle order and the main orientation order. Firstly, the color distribution features of interesting points are extracted and indexing is embedded into them, and tben the potential false matchings are discarded by the Kullback-Leibler divergence of corresponding color features. Secondly, the geometric loss score is computed according to the inconsistency between orders of the distance from points to the center of bundle region or the orders of the angle between the orientation of points and main orientation of bundle. The experiment on Copydays dataset shows the improved retrieval performance by adding color features and better circle order and main orientation order geometric constraints of the bundle feature in the image retrieval task.