针对当前广泛应用的BOVW模型存在精度不足问题,提出一种基于有序视觉词袋模型的相似性衡量方法.首先,对经过K-mean聚类得到的高维视觉单词,采用LLE(locally linear embedding)流形学习算法降至一维,对一维数据进行排序,并以此顺序对高维单词排序获得有序词袋库;其次,对样本图像的所有局部特征,以该特征在词袋中对应的有序单词索引号构建图像局部特征谱;最后,对训练样本和测试样本的局部特征谱作差求得残差,并以残差的1-范数衡量图像的相似性.KITTI数据集相似性衡量实验表明,有序BOVW模型相似性识别率明显高于无序BOVW模型.
In view of the shortcomings of the BOVW(bag of visual words) model widely used at present,a similarity measurement method based on ordered bag of words model was proposed.Firstly,the LLE(locally linear embedding) manifold learning algorithm was used to reduce the high dimensional visual words obtained by K-means clustering to one dimension.Then the one-dimensional data was sorted,and the high-dimensional words were sorted in this order to get the ordered bag of words library.Secondly,the local features of each sub image block were computed for all sample images.And the corresponding indexes in the word bag of the local features were utilized to construct the local feature spectrum of the images.At last,the residuals were calculated by subtracting the feature spectrum between the training and testing samples,and the similarity was measured by 1-norm of residuals.The similarity test results of KITTI data set show that the recognition rate of ordered BOVW model is significantly higher than that of disordered BOVW model.
Journal of Huazhong University of Science and Technology(Natural Science Edition)
ordered bag of words model
local feature spectrum
data dimensionality reduction