Finding and discriminating faces using biological barcodes

Citation

Dakin SC & Watt R (2009) Finding and discriminating faces using biological barcodes. In: Stoica A, Arlsan T, Erdogan A, Higuchi T, Bouridane A & El-Rayis A (eds.) 2009 International Symposium on Bio-inspired, Learning, and Intelligent Systems for Security, BLISS 2009. Los Alamitos, CA, USA: IEEE Computer Society Press, p. 50. http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=5376852; https://doi.org/10.1109/BLISS.2009.26

Abstract
We have recently presented evidence that faces elicit a unique pattern of response from filters that emulate the operation of neurons in the primary visual cortex [1]. Specifically, filters tuned to horizontal orientations produce a pattern of response that is reminiscent of a "barcode": a series of vertically co-aligned horizontal stripes of various heights and alternating contrast polarity. We review the evidence that this representation qualitatively explains many aspects of face perception (including the disruptive effects of spatial inversion and photographic negation) and then consider the computational advantages of what is essentially a onedimensional code. We also present evidence suggesting that barcodes support (a) discrimination of faces from one another and (b) face localisation. In the first experiment we measured subjects probability of classifying a (morphed) mixture of two faces, A and B, as being more like A or B. We can then derive a psychometric function by plotting the probability that subjects classify a face as being more like A, as a function of the ratio of A:B present in the morph. The slope of this function is an estimate of the mutual discriminability of faces A and B and, we show, is determined predominantly by horizontal information in the face. In the second experiment we presented faces embedded in natural scenes and measured the minimum time required to locate the face 75% of the time (the threshold exposure duration). Subjects were systematically faster at finding faces within horizontally filtered images than for other orientations. Our results indicate that humans rely heavily on horizontal information to find and discriminate between faces.)

Keywords
Face localization; Face recognition; Spatial filtering