A Comparative Study of Distance Measures for 2DPCA-Based Face Recognition

Face recognition is considered a primary technology in biometric systems. Despite its merits, the systems suffer because of the various facial variations. Two-Dimensional Principal Component Analysis (2DPCA) limits these variations by preserving facial spatial information and, at the same time, improves the computational efficiency. Those features give it an advantage in different applications, not only in face recognition but also in artificial intelligence (AI) applications. However, the distance methods used for classifications associated with 2DPCA are a significant issue. This paper explores and modifies various distance measures defined in the literature of the PCA approach for face recognition. These fourteen-distance metrics are modified and compared to the standard Euclidean distance on well-known face databases, namely the ORL, AR, and LFW databases. The experiments on these databases, which have diverse variations in facial images, manifest the superiority of the correlation coefficient-based distance method. Additionally, the results also show remarkable performance of angle-based and Canberra distance methods. These results clarify the importance of this work in enhancing distance metrics within 2DPCA algorithms to enhance the accuracy and robustness of both traditional and AI-driven recognition systems.