Johnson's Criteria for Pixel Resolution
Four Levels of Discrimination¹

Johnson's criteria  ±50% defines the number of picture elements, pixels, needed for each level of discrimination

(1) Detection - an object is present: 2 +1/-0.5 pixels
(2) Orientation - symmetrical, asymmetric, horizontal or vertical: 2.8 +0.8/-0.4 pixels  
(3) Recognition - the type object can be discerned, a person vs. a car: 8 +1.6/-0.4 pixels
(4) Identification - a specific object can be discerned, a woman vs. a man, the specific car: 12.8 +3.2/-2.8 pixels

                Commentary

These numbers come out of a military context, but are applicable in general.
Consider the following AT BEST - highly optimistic

(1) To detect a head in the image (on an otherwise featureless background since a non-featureless background introduce clutter that complicates detection) takes a minimum of ≈2 pixels wide, ≈3 pixels high, or an area of ≈6 pixels. 
(2) To orientate vertical rather than horizontal takes a minimum of ≈3 pixels width, ≈4 pixels high, or an area of ≈12 pixels. 
(3) To recognize a head rather than a tree, takes a minimum of ≈8 pixels width, ≈10 pixels high, or an area of ≈80 pixels. 
(4) To identify a  head takes a minimum of ≈13 pixels wide, ≈15 pixels high, or an area of ≈195 pixels. 
These are most likely optimistic for the following reasons.

Facial features, without relatively harsh, probably directional illumination tend to be relatively low contrast, thus suggesting that these numbers are overly optimistic since they require "adequate" contrast. Adequate S/N (signal to noise) would suggest the image was recorded with good lighting and a clean signal is available. A digitized image from a surveillance camera can be considered to provide an adequate S/N at ≈6 pixels area level. This again suggests the numbers above are overly optimistic.

Image processing can enhance both contrast and S/N, but , if the information isn't recorded, you can't synthesize it from nothing. Image processing can enhance what's there, but it adds nothing that is real. For example, if the grayscale range of the image is say 50 to 60, you may not be able to see much, but if you stretch the histogram to 10 to 240, you can certainly see the structure. If it is noisy, you can reduce the noise, but without enough pixels to be able to do some averaging of some kind, there isn't much gained - more likely you lose. 

If you have a reference image - a previous image of the face you are trying to identify - and if it happens to be in about the same orientation and scalable to the same size, you might be able to compare the known and unknown and get some kind of confidence level of it being a match, but I'm afraid the confidence would be pretty low at the 6 pixel level.

Bottom line - I wouldn't count on recognizing faces from 6 pixels. Astronomers might be able to claim there is a star in images of the sky where the star is only a few pixels across, but that's a point of light on a contrasting background - it's not recognizing a face, and it's more often than not differentiated from noise by comparing multiple images. Suspects in surveillance videos aren't exactly cooperative in repeatedly dropping by for a series of shots.  

A six pixel image is hopeless.  The image processing and analysis you can do is only as good as the information that you begin with. In other words if it is not there it is just not there. I love how Hollywood is beginning to give people the impression that image analysis is a cure all, and can make something from nothing. 

Copyright © 1998-2007 TREC  All Rights Reserved    Last Updated Thursday, May 08, 2008