多层融合深度局部PCA子空间稀疏优化特征提取模型

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.

Feature Extraction Model Based on Multi-layered Deep Local Subspace Sparse Optimization

Subspace method is classical pattem recognition method,that uses global information mainly to denote an image.Recently,with the introduction of deep learning,the feature extraction model based on local self-learning has attracted more and more attention.By using the theory of deep learning,this paper presents a new feature extraction model based on multi-layered deep local subspace sparse optimization to solve the problem of object recognition.Firstly,we calculate the PCA mapping matrix on the first layer by minimizing the reconstruction error on the training sample set,then we optimize the feature mapping results through L1 norm to enhance the robustness of algorithm.Secondly,we use the output of the first layer as the input of second layer,then we implement same actions of feature learning.In this way we can map the image to deep PCA subspace.Finally we merge these feature extraction results from different layers with weighting and encode the merged feature with binary hash code and histogram segment code.After that,we obtain the multi-layered deep local subspace sparse feature.The experimental results on face database of FERET 、AR 、Yale and target database of MNIST 、CIFAR-10 show that this feature extraction model can achieve high recognition rate and robustness for illumination,expression and pose.At the same time,compared with the convolutional neural networks,our algorithm owns the advantages of simple structure and fast convergent rate.