Stacked partial least squares regression for image classification

In recent years, the researches based on Convolutional Neural Network (CNN) have been doing in computer vision after the success in ILSVRC 2012. Hierarchical feature extraction is one of the reasons why CNN gives the state-of-the-art performance. On the other hand, Partial Least Squares (PLS) Regression which has been widely used in chemo-metrics is also used in computer vision in recent years. If class labels are used as objective variables for PLS, PLS can extract features suitable for classification. In this paper, we combine the idea of hierarchical feature extraction of CNN with feature extraction suitable for classification by PLS and propose a new method called Stacked PLS. It extracts features hierarchically in reference to CNN using PLS. The proposed method is evaluated on the MNIST dataset. Our method gave higher performance than CNN with the same network architecture and is comparable to the state-of-the-art methods.