一种基于词袋模型的大规模图像层次化分组算法

Large-scale image hierarchical grouping algorithm based on bag-of-words

大规模图像集合的自动分组,不仅可以帮助用户快速组织和掌握图像集合的内容,并且是基于图像的三维场景重建应用的前提和重要环节。提出一种基于词袋模型(bag-of-words,BOW)的层次化分组算法,将每幅图像表示为一个超高维视词向量,利用多路量化技术将内容相似的图像量化到同一个节点,从而完成对图像粗略分组。然后,在每组类别里面,对图像的局部特征向量进行逐一匹配,并利用仿射空间不变量的约束条件,去除不可靠特征匹配,得到更为准确可靠的图像相似度度量,从而完成图像的精细分组。实验结果表明:从得到的系统不同阶段图像分组的查准率-查全率(precision-recall)曲线可以看出,精细分组过程可以显著提高粗分组精度,并且在精细分组阶段。

Automatical grouping algorithm on large-scale image set,which is an important part of the scene reconstruction system,can help users organize the image set contents quickly.A hierarchical image grouping algorithm based on bag-of-words(BOW)was proposed.Firstly,each image is projected to a super-high dimensional visual word vector by a multiple paths quantization(MPQ)method,and this step is so-called coarse grouping.Then,feature matching is carried out in every divided group and an affine invariant constraint is proposed to get rid of the incorrect matching features.This step is so-called refined grouping which can improve image grouping accuracy.The precision-recall curves show that the refined grouping can obviously improve the accucy of coase grouping,and better grouping accucy can be achieved when using constraints.