Evaluating a Convolutional Neural Network on Short-Wave Infra-Red Images

Machine learning algorithms, both traditional and neuralnetwork-based, have been tested against RGB facial images for years, but these algorithms are prone to fail when illumination conditions are insufficient, for example, at night or when images are taken from long distances. Short-Wave Infra-Red (SWIR) illumination provides a much higher intensity and a much more ambient structure than visible light, which makes it better suited for face recognition in different conditions. However, current neural networks require lots of training data, which is not available in the SWIR domain. In this paper, we examine the ability of a convolutional neural network, specifically, the VGG Face network, which was trained on visible spectrum images, to work on SWIR images. Utilizing a dataset containing both RGB and SWIR images, we hypothesize that the VGG Face network will perform well both on facial images taken in RGB and SWIR wavelengths. We expect that the features extracted with VGG Face are independent of the actual wavelengths that the images were taken with. Thus, face recognition with VGG Face is possible between the RGB and SWIR domains. We find that VGG Face performs reasonable on some of the SWIR wavelengths. We can almost reach the same recognition performance when using composite images built from three SWIR wavelengths probing on RGB.