EVALUATING SHORT-WAVE INFRARED IMAGES ON A CONVOLUTIONAL NEURAL NETWORK
Michael Bihn · Digital Collections of Colorado (Colorado State University) · 2017
Machine learning algorithms, both neural networks and non neural networks, have been tested against RGB facial images for years.Their success is dependent on clearly distinguishable cropped faces.Poor lighting and distance are not beneficial.Short-wave infra-red (SWIR) imaging is higher in intensity and deeper penetrating than visible light.SWIR imaging overcomes the shortcomings of visible light.In adverse conditions SWIR can provide the images that visible light cannot.If we had lots of SWIR images we could train the neural networks as we do with the visual spectrum (RGB), but the data is expensive to produce.In this paper we examine the proficiency of a convolutional neural network facial recognition tool's ability to process SWIR images.We utilized a dataset containing both RGB and SWIR images.Our hypothesis is that VGGFace, trained on RGB images, will perform as well on SWIR 1300nm facial images as it does on RGB images.We expect VGGFace's feature extraction will work as well on SWIR images as RGB images, if not better.VGGFace performed better on the RGB images than the images captured in the SWIR 1300nm wavelength.The data failed to support our hypothesis while also revealing unexpected performance on the other SWIR wavelengths (935nm, 1060nm, and 1550nm).iii Remove the Profile Probe on Profiles . . . . . .