APP下载

一种快速低秩的判别子字典学习算法及图像分类

2021-07-11赵雅王顺政吕文涛王成群

智能计算机与应用 2021年1期

赵雅 王顺政 吕文涛 王成群

摘 要:本文提出了一种快速低秩的判别子字典学习算法。在训练阶段,构造一个子字典的低秩约束项和拉普拉斯矩阵正则化项,加入判别字典学习的目标函数中。将原始样本映射到一个新的空间中,使同一类别的相邻点彼此靠近,同时增强子字典对同类样本的重构能力,针对每类样本的判别性特征,学习出相应的学习字典。在测试阶段,利用kNN分类器估计测试样本的类别标签。同时,将算法应用在3种数据集上,与其他的字典学习算法进行比较,取得了较好的分类结果。

关键词: 子字典;判别字典;拉普拉斯矩阵;图像分类

文章编号: 2095-2163(2021)01-0051-04 中图分类号:TP391 文献标志码:A

【Abstract】This paper proposes a fast, low-rank discriminative sub-dictionary learning algorithm. In the training phase, the low-rank constraint terms of the sub-dictionary and the Laplacian matrix regularization terms are constructed, and the objective function of the discriminative dictionary learning is added. The original sample is mapped to the new space so that adjacent points of the same category are closed to each other. At the same time, the sub-dictionary's ability is enhanced to reconstruct similar samples, and the corresponding learning dictionary is learnt according to the discriminative characteristics of each sample. In the testing phase, the kNN classifier is used to estimate the class label of the test sample. Finally, the algorithm are applied to three public data sets compare with other dictionary learning algorithms. The proposed algorithm has achieved better classification results.

【Key words】sub-dictionary; discriminant dictionary; Laplacian matrix; image classification

0 引 言

判别字典学习是稀疏表示问题的一个研究分支,主要是通过重构训练样本得到样本的学习字典,并通过构造不同的约束项模型来增加字典的判别性能。字典学习包括无监督字典学习和有监督字典学习。其中,无监督字典学习主要是通过所有训练信号重建并优化字典,而不给出任何标签信息。典型的无监督字典学习有KSVD算法[1],MOD算法[2]等。Zheng等人[3]给出了使用拉普拉斯算子的无监督字典的图形正则化稀疏编码,并验证了其在分类和聚类上的有效性。但该类算法无法有效利用样本的标签信息,分类性能不一定是最佳的。相应地,有监督字典学习根据训练样本的标签信息学习出判别性字典。例如文献[4-7],就是典型的有监督学习字典。文献[5]提出了一个LSDDL算法,针对样本的局部特征和几何结构,结合样本的标签信息进行字典学习。……

登录APP查看全文