Hyperspectral Face Recognition via Existing 2D Face Recognition Methods

Hyperspectral imaging provides new chances for face recognition by improved discrimination in the spectral dimension. In this paper, we investigate hyperspectral face recognition by studying twenty-four existing 2D face recognition methods in combination with collaborate representation-based classifier (CRC). Since hyperspectral face data cubes contain significant amount of noise, we perform denoising on hyperspectral face data cubes by means of minimum noise fraction and video block matching 3D filtering. We crop the face images by a bounding box and use this bounding box image to classify the testing faces. We also conduct voting from the CRC output to decide the label of the unknown face cubes. Experiments show that histogram of oriented gradients achieves 100% correct classification rate for the CMU-HSFD dataset and local binary patterns obtain 96.8% correct classification rate for the PolyU-HSFD dataset.