Unsupervised Two-Stage TR-PCANet Deep Network For Unconstrained Ear Identification

Abstract

The main aim of this paper is to present a novel speedy, lightweight, and efficient two-stage TR-PCANet model for features extraction. TR-PCANet uses PCA to learn the filters of the convolutional layers. In order to generate powerful filters, we propose to augment the data used for the training. The Filter Learning stage is followed by binary Hashing and Blockwise Histogramming stages. At the end of the network, we propose normalizing the histograms using Tied Rank normalization. Moreover, as it could positively affect the identification rates, we suggest reshaping all images using CNN as a preprocessing stage. To further enhance the recognition yields, we combine TR-PCANet with TR-ICANet and DCTNet. We conduct extensive experiments on the public AWE dataset. The obtained results have proven the efficiency of the proposed network against TR-ICANet and DCTNet as well as the relevant state-of-the-art methods including deep learning-based ones.