Effect of Number of Face Images based on Illumination Variation in the Training Process on Face Recognition Results

Abstract

The research is related to face recognition which is influenced by illumination factor. The method used is the Robust Regression, which has a better performance than many other methods. The empirical experiment, which uses Yale Face Database B Cropped, is conducted to determine the effect of number of face images in the training process on face recognition perfomance. The hypothesis proposed in this research is the greater number of face images will result in higher facial recognition performance. The empirical experiment was conducted on this research to prove the hypothesis. Based on experiments that have been done, in general, the process of data training with many images will result in high performance of face recognition. But, this trend only occurs in images in the similar illumination condition. Illumination variation of face images also have significant impact on face recognition results. The process of training data with images of illumination variations (from several subsets of the face database) results in better face recognition performance than the process of training data with images of similar illumination conditions (from a subset of the face database). By using 19 images from subset 5 of the face database, face recognition accuracy is obtained at 95.11%. Whereas by only using 5 images from several subsets, obtained face recognition accuracy up to 96.10%. Even by using 7 images from several subsets, the accuracy obtained is up to 99.47%.

Keywords: Face Recognition Performance, Robust Regression, Data Training