Illumination Invariant Efficient Face Recognition Using a Single Training Image

This paper presents a single sample face recognition technique which takes care of illumination variations by applying normalization based on Weber's law. Local Directional Pattern (LDP) features are extracted from the normalized face by examining the prominent edge directions at each pixel. The LDP image is divided into non-overlapping windows and each window is treated as a fuzzy set. Treating LDP values as the information source values, entropy features called the information set-based features are extracted from each window. Further, 2DPCA is used to reduce the number of features. These features are augmented with entropy features of the fiducial regions and contour based features for face recognition. A nearest neighbor classifier based on these features is used on Extended Yale B and Face94 datasets and it is shown that compared with other results based on single and multiple training images, the proposed approach results in better recognition accuracy for wide illumination variations in test images. Further the efficiency of the scheme is shown by comparing the number of features needed for recognition.