PCA Dimensionality Reduction Method for Image Classification

The pooling layer has achieved good results in reducing the feature dimension and parameters of convolution neural network (CNN), but it will cause different degrees of information loss. In order to retain as much feature information as possible, we design a pooling method based on Principal Component Analysis (PCA)-P(CA)Pool. Firstly, all feature maps are traversed with the pooling window in which the data is extracted and stretched into row vectors. With the sliding of the pooling window, all row vectors are arranged in the matrix to form the sample matrix. Then all eigenvectors are extracted from the sample matrix by PCA algorithm to form the eigenvector matrix, which right multiplies the sample matrix to get the principal component matrix. Thirdly, each column of the principal component matrix is weighted with information coefficient which is determined by training to get the pooling vector. Finally, P(CA)Pool result is obtained by blocks arrangement of pooling vector. P(CA)Pool is tested with CNN-Quick, NIN, WRN-SAM, GP-CNN on datasets MNIST, CIFAR10/100 and SVHN. We also used AlexNet on Imagenet2012 to test P(CA)Pool. The experiment results show that compared with traditional pooling methods, P(CA)Pool could retain information in the pooling window better and improve the image classification accuracy.