组合金字塔和多核学习的图像分类方法  

Image Classification Using Spatial Pyramid and Multiple Kernel Learning

作　　者：甘玲[1] 谷伟庆 

机构地区：[1]重庆邮电大学计算智能重庆市重点实验室,重庆400065

出　　处：《小型微型计算机系统》2014年第7期1642-1646,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61075019)资助

摘　　要：针对图像分类空间语义信息贫乏,提出一种基于空间金字塔的多核学习图像分类方法.从图像中提取局部特征DenseSIFT,使用局部约束线性编码方法对这些特征进行编码;为增加图像分类时的空间信息以及图像块在图像分类时所起作用不同,采用空间金字塔多划分加权方式为图像加入空间信息;在图像分类时,采用多核学习方法为空间金字塔每一层的每个图像块各自形成一个核矩阵,学习各个核矩阵的权值,通过广义多核学习来获得区分能力最强的核矩阵.通过在Scene categories数据集上的实验,证明空间金字塔每一层的每一个图像块对图像分类集的区分能力不同,这也证明了所提出方法的有效性和鲁棒性.A novel image classification method based on spatial pyramid and multiple kernel learning is proposed in the paper. First,Dense-SIFT are extracted from images. Then,Locality-constrained Linear coding is adopted to represent an image as a vector. To add the spatial information,and each image blocks has different weights,We add this spatial information by dividing the image with the spatial pyramid. In the image classification,We use the multiple kernel learning,each blocks of spatial pyramid has its own kernels,and we adopt machine learning for the optimal trade-off between different kernels. A much more discriminative kernels can be seen as the generality multiple kernel learning of base kernels corresponding to different pyramid blocks. The experiment results conducted on the Scene categories show that each blocks of the spatial pyramid has different discriminative power,The experiments on the benchmark dataset show the effectiveness and robustness of our method.

关 键 词：图像分类 多核学习 空间金字塔 局部约束线性编码 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]