改进K-Means聚类算法及其在视觉词典构造中的应用

IMPROVED K-MEANS CLUSTERING ALGORITHM AND ITS APPLICATION IN VISUAL VOCABULARY BUILDING

K-Means聚类是视觉词典构造的常用方法,其聚类结果直接影响后续的特征量化效果和检索精度,而现有的K-Means聚类算法难以获得高质量的视觉词典。针对这种情况,提出局部化K-Means聚类算法。算法首先根据启发式原则将特征集划分成若干个独立的子集,并对各子集进行传统K-Means聚类,然后以各子集的聚类中心为对象进行加权K-Means聚类。上述过程不断迭代直至形成特定规模的视觉词典。实验结果表明,与现有算法相比,该算法提高了聚类质量。在SIFT特征集和标准数据集上进行的多组对比实验证明了该算法的有效性。

K-means clustering is a widely used method in visual vocabulary building,and its results will directly affect the subsequent quantisation quality and retrieval precision.However,the visual vocabulary obtained by existing k-means clustering algorithms is hardly to be high-quality.In view of this,we propose a localised k-means clustering algorithm.In the algorithm,firstly the feature set is heuristically di-vided into several independent subsets,and each subset is applied the traditional k-means clustering,then the weighted k-means clustering is executed on cluster centres of each subset.The above steps are iterated incessantly until the visual vocabulary with special size built.Experi-mental results show that this algorithm improves the clustering quality compared with existing algorithms.Multiple sets of comparative experi-ments conducted on SIFT feature set and standard dataset prove the effectiveness of the algorithm.