On transforming statistical models for non-frontal face verification
Pattern Recognition
| Hi-index | 0.00 |
It has been recently shown that local feature approaches to face verification are considerably more robust than holistic approaches, in terms of translations (caused by automatic face localization) and pose variations. In this paper we first investigate whether features based on local Principal Component Analysis (LPCA) are more discriminative than features based on the 2D Discrete Cosine Transform (2D DCT). We also investigate several methods for modifying the two feature extraction techniques in order to counteract the effects of linear and non-linear illumination changes, without losing discriminative information. Results on the XM2VTS database show that when using a Bayesian classifier based on Gaussian Mixture Models (GMMs), the performances of 2D DCT and LPCA techniques are quite similar, suggesting that the 2D DCT technique is preferable due to its lower computational complexity. When using 8脳8 blocks, modifying the 2D DCT and LPCA techniques by removing the first coefficient, which is the most affected by illumination changes, enhances robustness with little change in discrimination ability; removing further coefficients causes a noticeable reduction in performance on clean images and provides little gain in robustness. When using the 2D DCT with 16脳16 blocks, the first three coefficients need to be removed in order to achieve good robustness. It is further shown that contrary to previously published results, the use of deltas of low-order coefficients (to alleviate performance losses caused by removing coefficients) can adversely affect robustness.