摘要
提出基于类词包(CSBOW,Class-Specific Bag of Words)技术的有监督图像分类算法。在训练阶段,首先在每一类训练图像库中提取特征点,运用SIFT描述符形成特征区域描述符,然后根据欧氏距离对特征矢量进行K均值聚类,每一类形成若干聚类中心即所谓类词包。在测试阶段,首先提取测试图像的特征点,运用SIFT描述符形成若干特征矢量,然后依次计算每个特征矢量与各聚类中心的欧式距离,运用竞争投票机制决定图像的分类结果。将CSBOW技术应用到商品图像分类,实验结果证明本算法分类正确率高于传统的基于词包和SVM的图像分类算法。
A new supervised image classification algorithm based on CSBOW(class-specific bag of words) was proposed to have keypoints extracted from each image category in the training phase,then to generate characteristic area descriptors with the SIFT descriptor and to cluster the descriptors of each category with Kmeans so as to construct class specific words;in the test phrase,it has keypoints of test image extracted and represented as SIFT descriptors,then has the Euclidean distances from these SIFT descriptors to the words of each class calculated and the classification result achieved through the competitive voting.Experimental results prove the better performance of this proposed method compared to the traditional BOW with SVM.
出处
《化工自动化及仪表》
CAS
2012年第11期1465-1467,1525,共4页
Control and Instruments in Chemical Industry
基金
国家中小企业创新基金资助课题(09c26222123243)
关键词
图像分类
类词包
竞争投票
image classification
CSBOW
competitive voting