【作者】 刘扬； 程健； 卢汉清；

【Author】 LIU YANG CHENG JIAN LU HANQING(National laboratory of pattern recognition,Institute of automation,Chinese Academy of Sciences,,Beijing,100190)

【机构】 中国科学院自动化研究所模式识别国家重点实验室；

【摘要】 区别于传统的图像分类方法,本文利用迁移式学习的方法,通过从大量未标记且与待分类图像无关的图像样本上提取稀疏性局部特征来增强图像分类的效率。文中采用稀疏编码来得到待分类图像的紧致输入表示,但是一些关键步骤的改进使得我们的算法具有更好的效果。首先,当使用稀疏编码从未标记样本中提取图像块来学习基向量时,我们使用 SIFT 特征点来代替随机挑选点,使得学习得到的基向量更加能反映图像的局部结构。另外,我们发现二维 SIFT 特征空间相比于图像的灰度空间是一个更高阶的特征子空间,包含更多的局部信息。为了证实我们的理论推断,我们在 Caltech-101和 PASCAL VOC06数据库上进行了实验,和一些相关的图像分类算法的实验比较, 并将我们提出的算法结合 PMK 核空间理论,实验表明我们的算法具有更好的分类效率。更多还原

【Abstract】 Differently from traditional image classification algorithms,we apply transfer learning to address the problem of image classification with unlabeled and irrelevant images by learning more distinctive sparse local feature representations.In this paper,inspired by Raina’s work,we also use sparse coding to obtain the succinct input representations,some key improvements make the proposed algorithm more effective and efficient than Raina’s work.First,when choosing the patch from unlabeled image to learn the bases,we apply SIFT points rather than select points randomly,which makes the bases more meaningful.Second,we find that SIFT 2-D feature space is more exact high-level projections subspace than pixel intensity space to learn the bases.To validate our theoretic conjecture,we conduct the experiments on Caltech-101 data sets and PASCAL VOC06 data sets,we compare the proposed algorithm with some related work,and combine with PMK.The experimental results shown that the proposed algorithm obtains the significant improvement on the classification accuracy.更多还原