A Study on Optimal Face Ratio for Recognition Using Part-Based Feature Extractor

This paper aims to investigate the optimal face ratio for recognition. Face data are normalized to several ratios, which are 25%, 50% (equivalent to right and left face), and 75% of the full-face. The advantages of using different face ratios are these face data reduce the amount of computational power and storage requirements significantly. For fair comparison, various part-based linear subspace feature extractors, namely Non-negative matrix factorization (NMF), Local NMF (LNMF) and Spatially Confined NMF (SFNMF) are used to estimate the optimal face ratio. Our results show that 75% faces are good enough to produce demonstrably recognition accuracy.