Human face verification under illumination variation

Abstract

The appearance of a face will vary intensely when the illumination changes. The changes in the illumination conditions during image capturing make it difficult to obtain accurate face verification. Changes in illuminations results in two main problems, which are reflections and shadows. One of the most important aspects influencing the verification accuracy is illumination normalization. This thesis explored the use of fusion normalization methods to improve the performance of face verification under illumination variation. It has been shown that a single normalization technique is inadequate to solve the problems of illumination. In this study, several normalization methods such as Discrete Wavelet Transform, Discrete Cosine Transform, and Classified Appearance based Quotient Image were investigated for illumination normalization. A verification process was performed for each normalization technique and the outputs of the process, which were the likeness scores would be fused together to improve the final output. In the verification step, Principal Component Analysis was used to reduce the vector size of image and Linear Discriminant Analysis was used to extract discriminative information. In addition, un-trained fusion methods such as Max-Rule, Min-Rule, and Ave-Rule were used to get a unified decision with a reduced error rate. Besides that, fusion normalization methods were also used to solve all problems caused by illumination. The experiments were done on XM2VTS and Yale database B. The results of this research showed that the efficiency of Ave-Rule technique is better than other methods for XM2VTS, and the best fusion method for Yale database B is Min-Rule. To evaluate the techniques, the results have been compared with the outcomes of the fusion of each pair of the normalization methods as well as the results obtained from using other techniques. The comparison showed that the fusion of the three normalization techniques offered a better performance as compared to the fusion of two illumination normalization methods. Furthermore, the performance of face verification based on the fusion of the normalization methods was better in comparison to a single normalization technique.