基于软加权映射的局部聚类向量表示方法

Vector of locally aggregated descriptor based on soft assignment approach

基于特征码本的图像分类方法依赖于需要特征向量与聚类中心之间的映射,然而硬加权映射方法导致了相似的特征向量被映射为不同的聚类中心,从而降低了分类的查全率。为此提出一种基于软加权映射的局部聚类向量表示方法。该方法首先用k均值算法将特征向量聚类为k个聚类中心,采用最近邻算法寻找最接近的s个聚类中心,通过特征向量与聚类中心之间的相似度和邻近程度构建软加权映射的局部聚类向量,然后统计特征直方图,最后用主成分分析减少特征直方图维度。实验结果分析表明,相比较硬加权映射方法,文中方法提高了约5%的分类准确率。

The traditional bag-of-words image classification approaches are based on feature vectors mapping to clustering centers by hard assignment, which will cause vision similarly features vectors being mapped to different clustering centers. In this case, we pro- pose a novel vector of locally aggregated descriptor based on soft assignment approach. Firstly, we associate local features with s near- by cluster centers instead of its single nearest neighbor cluster depending on the distance between the features and the cell centers by u- sing k-means clustering algorithm. Then, we construct vector of locally aggregated descriptors by computing distances and similarity between feature vectors and clustering centers. Finally, we use PCA algorithm to reduce the dimension of feature histogram. The ex perimental results show that the proposed method can improve 5 % accuracy rate.