A New Methodology for Face Recognition under Varying Illumination Conditions

Abstract. Variations in illumination are one of major limiting factors for face recognition system performance. Here, a new photometric face normalisation method based on the local Discrete Cosine Transform is presented. The proposed method is compared with the best state of the art preprocessing technique producing equally good results. Although the ob-tained error rates in the tested database are very similar, we show that both methods classify the images different and their fusion improve the results. Based on this, a method which combines both preprocessing techniques is presented and it outperforms the results obtained with the methods working individually in face recognition under varying lighting conditions. Although the performance is degraded on good quality images, on which the best results are obtained when no preprocessing technique is applied. Then, a method to determine if a face image is affected or not by illumination is also presented in this work and it is used in a general framework for face recognition, in which the combination of preprocessing methods is only applied when the image is classified as bad illuminated while is directly classified if not, obtaining in this way very good results for both cases, permitting to apply this framework when the illumination conditions are variable and unknown