Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (12): 2374-2389.DOI: 10.3778/j.issn.1673-9418.2104109

• Graphics and Image • Previous Articles     Next Articles

Denoising Latent Subspace Based Subspace Learning for Image Classification

YANG Zhangjing, WANG Wenbo, HUANG Pu, ZHANG Fanlong   

  1. School of Information Engineering, Nanjing Audit University, Nanjing 211815, China
  • Online:2021-12-01 Published:2021-12-09

基于潜子空间去噪的子空间学习图像分类方法

杨章静,王文博,黄璞,张凡龙   

  1. 南京审计大学 信息工程学院,南京 211815

Abstract:

To solve the problem that the performance of discriminant least squares regression (DLSR) is not robust to image noise in image classification, a denoising latent subspace based subspace learning (DLSSL) image classification algorithm is proposed. This method is different from the existing classification algorithm based on regression in framework. It introduces a latent subspace in the visual space and label space, and improves the traditional one-step image classification framework to two-step, that is, noise reduction before classification. This method firstly extracts high-order features of data into latent subspace by incomplete autoencoder, then uses the high-order features for regression classification. At the same time, the distance between samples in the class is controlled by the group kernel norm constraint. The introduction of latent subspace brings more flexibility to the algorithm framework, alleviates the differences of dimensions and characteristics between visual space and label space, makes the incomplete autoencoder effective in noise reduction, and improves the robustness of classification algorithm. A number of comparison experiments are designed on the face, biometric, object and deep feature datasets. The experimental results show that the proposed algorithm has strong robustness to the noise in the image, and the obtained projection matrix is more discriminative. Compared with the related image classification algorithms, this algorithm has better performance and stronger universality. Thus it can be effectively applied to various image classification tasks.

Key words: autoencoder, subspace learning, low-rank, denoising, image classification

摘要:

针对判别最小二乘回归(DLSR)对图像噪声鲁棒性不佳的问题,提出一种基于潜子空间去噪的子空间学习图像分类方法(DLSSL)。该方法在架构上不同于现有基于回归的分类方法,其在视觉空间与标签空间中引入一个潜在子空间,将传统的图像分类框架改进为两步,即先降噪后分类。该方法先通过欠完备自编码将数据的高阶特征提取到潜在空间,再利用此高阶特征进行回归分类,同时辅以组核范数约束控制类内样本间距离。潜在子空间的引入为算法框架带来了更多灵活性,缓解了视觉空间与标签空间中数据维度与特性的差异,使得欠完备自编码可以有效地对数据进行降噪,提升了分类算法的鲁棒性。在人脸、生物指纹、物体和深度特征数据集上设计了多组对比实验,实验结果表明,算法对于图像中的噪声具有较强的鲁棒性,获得的投影矩阵具有良好的判别性,相比现有图像分类算法,性能更好、普适性更强,能有效地运用于各种图像分类任务。

关键词: 自编码器, 子空间学习, 低秩, 降噪, 图像分类