PLDANet: Reasonable Combination of PCA and LDA Convolutional Networks

Abstract

Integrating deep learning with traditional machine learning methods is an intriguing research direction. For example, PCANet and LDANet adopts Principal Component Analysis (PCA) and Fisher Linear Discriminant Analysis (LDA) to learn convolutional kernels separately. It is not reasonable to adopt LDA to learn filter kernels in each convolutional layer, local features of images from different classes may be similar, such as background areas. Therefore, it is meaningful to adopt LDA to learn filter kernels only when all the patches carry information from the whole image. However, to our knowledge, there are no existing works that study how to combine PCA and LDA to learn convolutional kernels to achieve the best performance. In this paper, we propose the convolutional coverage theory. Furthermore, we propose the PLDANet model which adopts PCA and LDA reasonably in different convolutional layers based on the coverage theory. The experimental study has shown the effectiveness of the proposed PLDANet model.