Study Of Feature Extraction Algorithm Based On Deep PCA

Feature extraction is a key step in image recognition and machine learning.Many existing algorithms extract features mainly from the pixel level information of images and they can not effectively represent the deep structure of data,such as PCA,LDA and so on.Deep learning is a kind of multi-hidden layer sensor which originated from artificial neural network,which simulates the cognitive mechanism of human brain.However,the network of deep learning is complex,and it needs many network parameters which need to be iteratively solved.To solve these problems,feature extraction algorithms based on deep PCA are proposed.The main contents are as follows:1.Deep PCA can not extract nonlinear features hidden in the data,which will result in poor classification performance.To solve this problem,Deep PCA-KPCA is proposed,which use PCA and KPCA to extract features on the first layer and the second layer respectively.Final features are obtained by cascading the features of each layer.In order to extract discriminant features,Deep PCA-LDA is proposed by using LDA instead of KPCA on the second layer.Deep PCA-LDA makes full use of the tag information of the data,which is beneficial to image classification.Experiments show that our algorithms can effectively improve the performance of image classification.2?The input images of PCANet are divided into blocks,which can retain some spatial information.But,for the extraction of deeper structure information of the image,the image block processing is not enough.To solve these problems,we propose EPCANet.EPCANet includes two convolution layers,a data processing layer and an output layer.On convolutional filter layers,Principal Component Analysis(PCA)is used to obtain the filter kernels instead of gradient descent method.Thus,tuning parameter process is avoided.On data processing layer,the output of the first layer and the original image are pixel staggered or subsampled,which can make the information richer.Output layer consists of simple binary hashing and block-wise histograms for nonlinear processing.Extensive experimental results in several databases show that our algorithms can effectively improve the performance of image classification.