PCA-LDANet: A Simple Feature Learning Method for Image Classification

In this paper, we propose a simple and effective feature learning architecture for image classification that is based on very basic data processing components: 1) principal component analysis (PCA); 2) linear discriminant analysis (LDA); and 3) binary hashing and blockwise histograms. In this architecture, the PCA is employed to reconstruct patches of input images, and the LDA is employed to learn filter banks. This is followed by simple binary hashing and blockwise histograms for indexing. This architecture is motivated by LDANet and PCANet, thus called the PCA LDA Network (PCA-LDANet). They have some similarities in their topologies. We have tested the PCA-LDANet on two visual datasets for different tasks, including the Facial Recognition Technology (FERET) dataset for face recognition; and MNIST dataset for hand-written digit recognition. To explore the properties and essence of these architectures, we just conduct experiments on the one-stage networks. It is enough to explain the issue properly. Experimental results show that the PCA-LDANet-1 outperforms both PCANet-1 and LDANet-1 on both datasets. The experimental results demonstrate the effectiveness and distinctiveness of the PCA-LDANet; and the important role of PCA patch reconstruction in the PCA-LDANet.