Comparing Illumination Normalization Methods for Face Recognition

The effects of image illumination changes is one of the major concerns in face recognition because it is significantly degrading recognition accuracy. This work presents a comparative evaluation of eight illumination normalization algorithms (Gamma, DoG, Homomorphic, Gradient faces, SQI, Steerable, Isotropic, and Anisotropic) combined with a unified pipeline of Gabor feature extraction, PCA dimensionality reduction, and four classifiers (Naïve Bayes, SVM, K-NN, and MLP). Practical study is conducted using the Yale B, AR, and CMU-PIE standard image databases.
The best recognition rates (RR) were achieved as follows: on AR, DoG and Isotropic filtering reached up to 98.42% (MLP); on CMU-PIE, DoG and Anisotropic filtering achieved 56.84% (K-NN); and on Extended Yale B, SQI, Steerable, and DoG achieved up to 94.21% (MLP).
The results demonstrate that local normalization methods consistently outperform global methods under severe illumination. Therefore, DoG and Isotropic filtering are recommended for moderate lighting conditions, while SQI, Steerable filtering, and DoG are more suitable for extreme illumination scenarios.