Collaborative representation for visible to band gender classification using multi-spectral imaging: Extensive evaluations by exploring 22 photometric normalization methods

Abstract

Facial biometric trait based on visible spectrum have been widely studied for gender classification in recent times due to its non-intrusive nature of image capture in a covert manner. Recently, with the advancement in sensing technology, multi-spectral imaging has gained significant attention in this direction. Although substantial studies related to facial gender classification can be seen in the literature independently based on visible and limited on multi-spectral imaging, the essence of gender classification based on visible to multi-spectral band is still an open and challenging problem due to varying spectral gap. In this paper, we present an extensive benchmarking study that employs 22 photometric normalization methods to bridge the gap between visible face images in the training set and facial band images in the testing set before performing the classification. Furthermore, for gender classification, we employed a robust Probabilistic Collaborative Representation Classifier (ProCRC) to learn the features from visible face images and then testing with the individual bands. We present the evaluation results based on visible and multi-spectral face database comprises of 82,650 sample face images belonging to six different illumination conditions. We present a benchmarking quantitative and qualitative average gender classification accuracy by employing 10-fold cross-validation approach independently across nine spectral bands, six different illumination conditions, and 22 photometric normalization methods. Furthermore, to present the essence of our proposed approach, we compare our evaluation results against Convolutional Neural Network (CNN) classifier. The experimental results show maximum gender classification accuracy of 95.3 plus or minus 2.3 percent with our proposed approach, outperforming the state-of-the-art CNN demonstrating the potential of using photometric normalization methods for the improved performance accuracy.