A Neural Network-based Face Recognition Method by Illumination Normalization & Edge Detection

摘要

Neural networks have been extensively studied for applications related to face recognition. We compare the performance benefits of two preprocessing techniques, wavelet transform-based illumination normalization and edge detection, for neural net-based identity recognition systems. The methods are tested on a multi-pose, multi illumination source dataset consisting of images of 39 people over 9 poses and under 64 illumination conditions. We find that the wavelet transform-based illumination normalization network correctly classifies all of the testing images. Due to difficulties with the computing environment, the Canny edge detection method fails to complete training, but intermediate results suggest that it will perform better than no preprocessing but not as well as the illumination normalization. We plan to let it continue training over spring break to be able to view the true testing results.