Abstract

The principal component analysis network (PCANet) is a simple stacking-based convolutional neural networks, it achieves good precision in target recognition tests. However, the principal component analysis (PCA) is limited when dealing with high dimensional data. In this paper, the simplified local preserving projection network (SLPPNet) is proposed to improve recognition performance. In SLPPNet, LPP-like algorithm is applied to extract features, which can maintain the spatial structure of high-dimensional data. Moreover, we changed the structure of the PCANet to obtain the filter kernels. The PCANet needs to use all training images to calculate a multistage filter bank, while the SLPPNet only needs to use one training image. Hence, the SLPPNet can greatly reduce the amount of calculation. We tested the SLPPNet extensively on various databases, such as Yale, ORL, AR and Extended Yale B datasets, and compared it with the PCANet. Experimental results show that the accuracy of SLPPNet outperforms PCANet and the SLPPNet spends less time. Especially in the occlusion databases, the recognition accuracy of SLPPNet enhances remarkably.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call