多层融合深度局部PCA子空间稀疏优化特征提取模型

发布时间：2018-04-23 05:26

本文选题：深度学习 + 多层融合　；参考：《电子学报》2017年10期

【摘要】：子空间方法是主要利用全局信息的经典模式识别方法,随着深度学习思想的引入,局部自学习结构特征模型得到大家的关注.利用深度学习原理,本文提出一种多层融合的深度局部子空间稀疏优化特征自学习抽取模型解决目标识别问题.首先,对训练样本集通过最小化重构误差得到第一层的主成分(Principal Component Analysis,PCA)特征映射矩阵;然后,通过L1范数约束对特征映射结果进行稀疏优化,提高算法鲁棒性.接着,在第二层映射层以第一层的特征输出为输入,进行同样的特征矩阵学习操作,最终将图像映射至深层PCA子空间;然后,对各个映射层的特征提取结果进行加权融合,进行二值化哈希编码和直方图分块编码,提取图像的深度子空间稀疏特征.在FERET、AR、Yale等经典人脸数据库以及MNIST、CIFAR-10等目标数据库上的实验结果表明,该算法可以取得较高的识别率以及较好的光照、表情、人脸朝向鲁棒性,并且相对于卷积神经网络等深度学习框架具有结构简洁、收敛速度快等优点.
[Abstract]:Subspace method is a classical pattern recognition method which mainly uses global information. With the introduction of the idea of deep learning, the local self-learning structural feature model has attracted much attention. Based on the principle of depth learning, this paper presents a multi-layer fusion model of sparse feature extraction in deep local subspace to solve the problem of target recognition. Firstly, the first level principal component Principal Component Analysis (PCA) feature mapping matrix is obtained for the training sample set by minimizing the reconstruction error, and then the algorithm robustness is improved by sparse optimization of the feature mapping results with L1 norm constraints. Then, at the second layer, the feature output of the first layer is taken as the input, and the same feature matrix learning operation is carried out. Finally, the image is mapped to the deep PCA subspace, and then the feature extraction results of each mapping layer are weighted and fused. Binary hash coding and histogram block coding are used to extract the sparse feature of the depth subspace of the image. The experimental results on the classical face databases such as Yale and MNIST CIFAR-10 show that the algorithm can achieve higher recognition rate, better illumination, expression and face orientation robustness. Compared with the convolutional neural network, it has the advantages of simple structure and fast convergence.
【作者单位】：燕山大学信息科学与工程学院;
【分类号】：TP391.41

【相似文献】