Critical Section

Human Pattern Recognition

Sunday,  06/22/03  10:26 AM

Among all the remarkable things that we humans can do, effortlessly, that takes huge amounts of effort to do any other way; our image pattern recognition is perhaps the most awesome.

Original Image
Original Image

Poor Compression
Poor Compression

Poor Difference

Good Compression
Good Compression

Good Difference

I was just struck by this in working with various compression schemes for images.  The best way for me to evaluate a compression routine is to load an image in Photoshop, then load an image which has been compressed and decompressed with the routine in question, and create a new image which is the "difference" between them.  I can then visually tell right away how well the compression routine worked.  This is much better and much more informative than writing a program to perform a detailed comparison between the two images.

Not only does this tell me quickly how well the routine worked - the degree of artifacts introduced by the compression - but it tells me what kind of artifacts are introduced.  Does the routine introduce false edges?  Does it attenuate areas of high contrast?  Does it smear detail?  Does it affect all color channels equally?  Does it have problems with very light areas, or very dark?  The only reason I can discern these things so easily is that my personal pattern recognizer is so good.  Human vision is amazing.  It can pick up edges easily, with no effort required.  It can detect subtle differences in contrast, across a very high range of intensity.  It can compare two regions of high-frequency detail and tell them apart.

If you look at the images to the right, you can easily see the "poor compression" is inferior to the "good compression".  Your pattern recognition is that good.  You might not realize why; but looking at the difference images, you have quantitative confirmation, as well as information about the types of artifacts (note all the little horizontal lines!)  Writing a program to compare the images and find these artifacts would be non-trivial.

As part of Aperio's approach to Automating Pathology, we've created a lot of amazing tools to perform pattern recognition on very large images.  These tools are much faster and much more quantitative than anything a human can do.  But their qualitative abilities are vastly inferior.  In fact, a big part of the technique we use is to automate the laborious pre-processing of images to optimize the use of human pattern recognition, just as "differencing" with Photoshop does.

About Me

Greatest Hits
Correlation vs. Causality
The Tyranny of Email
Unnatural Selection
Aperio's Mission = Automating Pathology
On Blame
Try, or Try Not
Books and Wine
Emergent Properties
God and Beauty
Moving Mount Fuji The Nest Rock 'n Roll
IQ and Populations
Are You a Bright?
Adding Value
The Joy of Craftsmanship
The Emperor's New Code
Toy Story
The Return of the King
Religion vs IQ
In the Wet
the big day
solving bongard problems
visiting Titan
unintelligent design
the nuclear option
estimating in meatspace
second gear
On the Persistence of Bad Design...
Texas chili cookoff
almost famous design and stochastic debugging
may I take your order?
universal healthcare
triple double
New Yorker covers
Death Rider! (da da dum)
how did I get here (Mt.Whitney)?
the Law of Significance
Holiday Inn
Daniel Jacoby's photographs
the first bird
Gödel Escher Bach: Birthday Cantatatata
Father's Day (in pictures)
your cat for my car
Jobsnotes of note
world population map
no joy in Baker
vote smart
exact nonsense
introducing eyesFinder
to space
where are the desktop apps?