东篱科研大数据发现系统（DRDS）

位置：成果数据库 > 期刊 > 期刊详情页

利用图像类标信息的自调式字典学习方法

ISSN号：0469-5097
期刊名称：南京大学学报(自然科学版)
时间：0
页码：-
分类：TP181[自动化与计算机技术—控制科学与工程;自动化与计算机技术—控制理论与控制工程]
作者机构：[1]南京大学计算机软件新技术国家重点实验室,南京210023
相关基金：教育部新世纪优秀人才计划（NCET-11-0213）; 国家自然科学基金（61273257,61321491,61035003）; 江苏省六大人才高峰项目（2013-XXRJ-018）
相关项目：基于云计算的海量数据挖掘

关键词：类标签, 自调学习, 字典学习, label sign, self-paced, dictionary learning

中文摘要：

字典学习是图像分类的关键研究问题之一.现有的字典学习方法大都假设所有训练样本同等重要.实际上,训练样本由于样本之间关联性作为一种＂隐藏属性＂是未知的,因此,训练样本的学习顺序也与学习效果密切相关.提出一种将自调学习机制融合于字典更新过程的新型字典学习方法,在字典学习中,学习的过程并不是一次处理所有训练样例,而是从简单的训练样例学起,通过迭代逐步扩展至整个训练数据集.针对自调式过程是一种无监督式的学习这一特点,融合类标机制,利用图像类标信息进行监督,得到一种更加高效的简单样本判别方法,从而提高学习过程中反复迭代的效率.在Caltech-101数据集上进行图像分类实验,并和其他几种字典学习算法进行了分析和比较,结果表明本文算法在字典表示以及分类效果上都取得了更好的效果.

英文摘要：

Image classification is an important research task in multimedia content analysis and processing.Dictionary learning is an important research task in the state-of-the-art image classification framework.Most existing dictionary learning approaches ignore the＂hidden＂information of samples and assign equal importance to all training samples,which in fact have different complexity in terms of sparse representation.In all of the advanced methods,a self-paced dictionary learning algorithm solves the problem and has a well efficiency.The algorithm uses the easy samples to train the dictionary first,and then iteratively introduces more complex samples in the remaining training procedure until the entire training data are all easy samples.Meanwhile,label information with each dictionary item（columns of the dictionary matrix）as a＂hidden＂information is used to enforce discriminability in sparse codes during dictionary learning process.In this paper,we propose a self-paced dictionary learning algorithm using the class label of training data,to find an efficient method to determine the easy samples,and improve the classification accuracy.Specifically,we add a label consistent information called＂discriminative sparse-code error＂into the objective function in finding easy samples procedure to improve the efficiency of self-paced method.Experimental results on benchmark datasets Caltech-101 show that our algorithm leads to better dictionary representation and classification performance than the baseline methods under the same learning conditions.

同期刊论文项目