A Comparative Study on Canonical Correlation Analysis-Based Multi-feature Fusion for Palmprint Recognition

Yihang Wu, Junlin Hu

Abstract: Contactless palmprint recognition provides high-accuracy and friendly experience for users without directly contacting the recognition device. Currently, many existing methods have shown relatively satisfying performance, but there are still several problems such as the limited patterns extracted by single feature extraction approach and the huge gap between hand-crafted feature-based approaches and deep learning feature-based approaches. To this end, in this paper, we make use of multiple palmprint features and exploit the benefits of hand-crafted features and deep features in a unified framework using Canonical Correlation Analysis (CCA) method, and present a comparative study on CCA-based multi-feature fusion for palmprint recognition. In the experiments, the best feature fusion scheme achieves 100% accuracy on Tongji palmprint dataset and shows good generalization ability on IITD and CASIA palmprint datasets. Extensive comparative experiments of different approaches on three palmprint datasets demonstrate the effectiveness of CCA-based multi-feature fusion method and the prospects of applying feature fusion techniques in palmprint recognition.