摘要
在基于视觉词包模型的图像分类方法中,Fisher向量编码是常用的图像表示方法之一.该方法利用每一个特征关于所有高斯子模型似然函数的梯度信息来构建图像表达.而在编码过程中,每一个特征都会被投影到所有的高斯子模型上并进行编码,同时子模型之间的内在差异也未被考虑,这些不足削弱了Fisher向量的表达能力.为此,提出一种基于k密集近邻算法的局部Fisher向量编码方法.在编码过程中该方法引入局部性约束原则,并利用图像特征空间中高斯子模型间的拓扑结构差异.在多个数据集上进行测试,结果表明改进方法能够有效提升分类的准确率.
For the image classification methods based on the bag-of-visual-words model,Fisher vector(FV)encoding is one of the popular image representation approaches.In this method,gradient information of the likelihood functions,which is achieved by fitting each feature with all Gaussian sub-models,is used to build image representation.However,in this encoding procedure,each feature is mapped to all of the Gaussian sub-models and encoded by them,and the inherent differences between these sub-models have not been considered.These drawbacks limit the representative ability of the Fisher vector.To solve these problems,a local Fisher vector encoding approach based on k-dense neighborhood(KDN)algorithm is proposed,which introduces the local constraint and utilizes the difference between the topological structures of the Gaussian sub-models.Experiments are conducted on several benchmark datasets,and the results demonstrate the effectiveness of the proposed method in improving the accuracy of the classification.
作者
冀治航
胡小鹏
杨博
田云云
王凡
JI Zhihang;HU Xiaopeng;YANG Bo;TIAN Yunyun;WANG Fan(School of Computer Science and Technology, Dalian University of Technology, Dalian 116024, China;School of Information Engineering, Henan University of Science and Technology, Luoyang 471023, China)
出处
《大连理工大学学报》
EI
CAS
CSCD
北大核心
2020年第4期411-419,共9页
Journal of Dalian University of Technology
基金
国家重大专项资助项目(2018YFA0704605)
“十三五”重大专项资助项目(2017ZX05064)。
关键词
视觉词包模型
图像分类
Fisher向量编码
k密集近邻算法
bag-of-visual-words model
image classification
Fisher vector encoding
k-dense neighborhood algorithm
作者简介
冀治航(1979-),男,博士生,E-mail:jizhihang2000@163.com;王凡(1975-),女,博士,副教授,E-mail:wangfan@dlut.edu.cn。
