Pokerface
 
Pokerface: Partial Order Keeping and Energy Repressing Method for Extreme Face Illumination Normalization
F. Juefei-Xu and M. Savvides
IEEE 7th International Conference on Biometrics: Theory, Applications and Systems (BTAS), 2015.
 
@INPROCEEDINGS{Felix_btas15_poker,
author={Juefei-Xu, F. and Savvides, M.},
booktitle={Biometrics: Theory, Applications and Systems (BTAS), 2015 IEEE Seventh International Conference on}, title={{Pokerface: Partial Order Keeping and Energy Repressing Method for Extreme Face Illumination Normalization}},
year={2015},
month={Sept},
pages={1-8},
}
 

The Pokerface is an illumination normalization technique for faces under extreme illumination conditions. It first aims at maximizing the minimum gap between adjacently-valued pixels while keeping the partial ordering of the pixels in the face image under extreme illumination condition, an intuitive effort based on order theory to unveil the underlying structure of a dark image. This optimization can be formulated as a feasibility search problem and can be efficiently solved by linear programming. It then smooths the intermediate representation by repressing the energy of the gradient map. The illumination normalized faces using our proposed Pokerface not only exhibit very high fidelity against neutrally illuminated face, but also allow for a significant improvement in face verification experiments using even the simplest classifier. Simultaneously achieving high level of faithfulness and expressiveness is very rare among other methods. These conclusions are drawn after benchmarking our algorithm against 22 prevailing illumination normalization techniques on both the CMU Multi-PIE database and Extended YaleB database that are widely adopted for face illumination problems.

 

The Pokerface is very different from any previous work, it is hard to group it into any category. The main idea behind the Pokerface is very intuitive and straight-forward. It aims at transforming a dark image patch to a bright (illumination normalized) one with distinguishable details by keeping the partial orders of the pixels. As depicted in the Figure above, for each of the four images shown in the first row, a letter “S” pattern is embedded with the rest of the pixels being pure black (intensity=0) except the top-left pixel being pure white (intensity=255). However, we can hardly discern the “S” pattern from the black background in the left-most image. The histograms for these four images are shown in the second row. As can be observed, when the “S” pattern pixels have small intensity gap (\epsilon) between adjacently-valued pixels, the pattern is not as salient as the pattern showing larger gaps between adjacently-valued pixels such as the right-most image. The goal of the Pokerface is exactly to maximize the minimum gap (\epsilon) between adjacently-valued pixels while preserving the partial ordering of the pixels. The latter is to guarantee that the local edge information is well preserved when normalizing a dark image to a bright one. We will formulate this optimization problem in the context of order theory in the next section.

The optimization behind Pokerface goes like this. We scan each local neighborhood of the dark image and retain the partial order characteristics of the center pixels and all its neighbors. The ordinal relationships are then binarized and stored in matrix A and b, indicating location of each comparison as well as the sign. The goal of the Pokerface is to maximize the minimum gap between the recovered bright face, while keeping the same partial ordering as in the dark face. Therefore, this problem can be solved using aforementioned special case of linear programming, where y is the bright face we are recovering, and \epsilon is the minimum gap we need to satisfy.