基于SVM的BoVW距离度量学习

SVM Based BoVW Distance Metric Learning

在视觉单词包模型(bag of visual words,BoVW)模型中,由于特征检测的不足、聚类算法的缺陷及视觉单词的量化误差,用BoVW模型产生的视觉词典中,存在视觉单词同义性和歧义性的问题,因此用BoVW计算图像距离时,效果不太理想。BoVW模型产生的词典规模巨大,学习一个普通矩阵需要的运算量难以接受。针对BoVW模型上述缺陷,文章提出了一种基于SVM的BoVW距离度量学习方法。该方法利用SVM训练一个将相似图像对与非相似图像对最大程度分离的超平面,得到计算词频直方图点积的权重矩阵。在Oxford图像集上的检索实验表明了该方法的有效性。

Due to the imperfect feature detection, cluster algorithm drawback, and quantization error existing in bag of visual words （BoVW） , the visual vocabulary produced by BoVW has the visual word synonymy and polysemy problem, therefore it is not satisfactory to measure image distance u- sing BoVW. Aiming at the above-mentioned BoVW defects, this paper presents an SVM based BoVW distance metric learning. This approach uses SVM to train a hyperplane which ean separate similar image pairs and dissimilar image pairs to a large degree, and get a weight for each dimension used to compute the visual words histogram point product. Experiments on Oxford image datasets demonstrate that the approach is effective.