摘要
针对目前需要大量实验方可获得视觉字典的不足,提出了一种一次既可获得合理的视觉字典方法。首先,采用尺度不变特征转换SIFT[1](Scale-invariant feature transform)局部描述子构建场景图像数据集的特征矩阵;其次,采用AP聚类算法对场景图像的特征矩阵进行聚类,获得聚类中心数,也就是合理的视觉字典容量,并结合K-means算法获得共现矩阵,再用PLSA算法构建概率模型,然后用SVM[2]进行分类得出正确率。最后,用该方法与传统的通过大量实验的获得合理的视觉容量的方法进行对比分析主题数K(PLSA的参数之一)对实验结果影响。
For a Iarge number of experiments needed to quaIify for the current Iack of visuaI dictionary,one can propose a rea-sonabIe method for visuaI dictionary.FirstIy,SIFT feature scenes of IocaI descriptors to buiId a matrix of image data sets. SecondIy,AP cIustering aIgorithm for image feature matrix scene cIustering obtain poIy cIass center number,which is a rea-sonabIe visuaI dictionary capacity,combined with the K-means aIgorithm to obtain the co-occurrence matrix,then PLSA aIgo-rithm to construct a probabiIity modeI,and then use SVM arrive at a correct cIassification rate.
出处
《工业控制计算机》
2015年第4期114-115,117,共3页
Industrial Control Computer
