Identification of faces obscured by noise: comparison of an artificial neural network with human observers
Kim T. Blackwell, Thomas P. Vogl, HANS P.DETTMAR, Michael Anthony Brown Sr., G.S. Barbour, Daniel L. Alkon · Journal of Experimental & Theoretical Artificial Intelligence · 1997
Face identification is easily accomplished by humans but is an exceptionally difficult task for machine vision algorithms. This report is the first to directly compare the face identification performance of humans with that of an artificialneural network,Dystal, using digitized images of the faces of eight individuals in an eight-alternative forced-response paradigm. The test images differed from the training images in facial expression, head tilt and rotation, amount and correlation of added noise, and the presence of a stocking mask in some of the images. The images were deliberately not preprocessed by a feature extraction algorithm to avoid confoundingthe performance of Dystal with the performance of the feature extraction algorithm.While human observers outperform Dystal at low noise levels, at high levels of correlated noise Dystal outperforms human observers, who score just above the chance level. The greater sensitivity to noise exhibited by human observers is attributed to the local feature extraction performed by human observers, but not by Dystal.