Handling illumination variation in face recognition using multiscale retinex

Face recognition has an important role in the age of information technology to improve comfort and security. Various face recognition algorithms are proven to be effective if done under a controlled environment. However, recognition performance will decrease significantly in the uncontrolled environment, for example caused by illumination variation. This paper discusses the approach to deal with varying illumination on face recognition with Multi-Scale Retinex method. To verify accuracy of this algorithm, it is used Principal Component Analysis (PCA) and Eulidean distance as algorithms to identify faces. In the experiments, Multi-Scale Retinex is also compared to other normalization methods, such as gamma correction, histogram equalization, Retinex, and Single-Scale Retinex. Based on several face recognition dataset, that is Extended Yale B, Faces95, and Grimace, it can be concluded that the Multi-Scale Retinex method with one scale (Single-Scale Retinex) can properly normalize the illumination on the face image in term of speed and performance.