一种基于GMM-聚类树的面向大规模服装图像检索的索引结构

An Index Framework for Large Scale Clothing Image Retrieval Based on GMM-cluster Tree

目前服装图像检索研究主要偏重于考察服装图像底层特征的表示方法,对服装图像数据集整体的特性关注较少.服装图像种类、款式、细节多种多样,传统方法在检索速度和准确率上已经不能满足用户的需求.为在日益庞大的服装图像集中快速准确地搜索需要的款式,提出一种GMM-聚类树索引方式.GMM-聚类树将不同细节的服装图像按照相似性逐步聚类,分别对应于不同的聚类树分支,避免人为指定聚类个数造成服装分类的错误.分别在小数据集和大数据集上测试其对服装图像检索的准确率和效率,试验结果表明通过自动确定聚类个数和GMM-聚类树的逐层分类,能够带来检索准确率和效率的双重提升.

The current image retrieval researches are focused on the low-level features of clothing images, while the characters of the whole clothing dataset are ignored. The clothing images have many classes, styles and details, and the dataset is growing in an amazing speed, which brings great challenges to the traditional retrieval methods in accuracy and efficiency. To address this problem, a new index framework named GMM-cluster tree was designed, which could classify the clothing ima- ges and save them into corresponding tree branches according to their classes, styles and details through hierarchical clustering, so as to avoid the wrong clothing classification due to artificial desig- nation of cluster numbers. The accuracy and efficiency of clothing image retrieval are tested respec- tively based on a small and a large dataset. The experiment results show that both accuracy and effi- ciency of the research can be improved through the automatic determination of cluster numbers and layer-by-layer classification of the GMM- clustering tree.