Browse Prior Art Database

Improved Methods for Querying Image Databases using Texture

IP.com Disclosure Number: IPCOM000104876D
Original Publication Date: 1993-Jun-01
Included in the Prior Art Database: 2005-Mar-19
Document File: 4 page(s) / 131K

Publishing Venue

IBM

Related People

Equitz, W: AUTHOR

Abstract

Described is an improved method for querying image databases using the computed texture measures of [*]. The Tamura texture measures compute three main texture features called "contrast", "coarseness", and "directionality". In their paper, they also describe several other measures, but they conclude that these three are the most important. These measures are defined only for gray-scale images, so color images are first converted to luminance (gray-level) values before applying these measures. The contrast measure is the simplest, and is just a function of the luminance histogram. Specifically, Contrast = sigma/(alpha sub 4)sup 1/4, where alpha sub 4 = mu sub 4 / sigma sup 4 (kurtosis), and &mu. and &sigma. are the mean and standard deviation (respectively) of the luminance distribution.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 48% of the total text.

Improved Methods for Querying Image Databases using Texture

      Described is an improved method for querying image databases
using the computed texture measures of [*].  The Tamura texture
measures compute three main texture features called "contrast",
"coarseness", and "directionality".  In their paper, they also
describe several other measures, but they conclude that these three
are the most important.  These measures are defined only for
gray-scale images, so color images are first converted to luminance
(gray-level) values before applying these measures.  The contrast
measure is the simplest, and is just a function of the luminance
histogram.  Specifically, Contrast = sigma/(alpha sub 4)sup 1/4,
where alpha sub 4 = mu sub 4 / sigma sup 4 (kurtosis), and &mu.  and
&sigma.  are the mean and standard deviation (respectively) of the
luminance distribution.  This is very easy and fast to compute.  It
provides a measure of the global contrast (bright/dark) in the image.

      The coarseness measure provides a gauge on the granularity of
the image.  It is basically a "size detector", and involves computing
a moving average at a variety of window sizes (1x1, 2x2,..., 32x32).
Then for each pixel, one computes delta sub 1, delta sub 2, ....,
delta sub 32, where delta sub i = max(delta sup h sub i, delta sup v
sub i).  The variable delta sup h sub i is the difference between the
moving averages of size ixi immediately to the left and immediately
to the right of the ixi moving average centered around the current
pixel.  The variable delta sup v sub i is similarly defined for
vertical neighbors.  So, for each pixel, one gets a graph something
like in Fig. 1.  Here, one wants to find the largest window that has
within a few percent of the maximum delta sub i In this example, the
best window size would be 8.  To calculate the overall coarseness of
an image, all one needs to do is to take the average best size over
all the pixels in the image.

      The problem with direct application of this method is that for
smaller images (or objects within an image) one doesn't have enough
pixels to calculate moving averages for the larger window sizes, and
the method is undefined if one doesn't have a delta sub i for every
&iota.  If one simply chooses the best window size from the available
delta sub i values, then small objects always get small coarseness
values and are classified as grainy objects, making the query results
not that great.

      So, the innovation necessary to make the Tamura coarseness
measure practical involves finding a way to deal with small objects.
The solution used in our implementation is to assume that the best
window size was the maximum available (32 X 32 in our example) when
for every i for which delta sub i is defined, delta sub i is within a
certain small percentage of maxsub j lt sub i delta sub j, where the
max (o) is performed over all j lt i.  That is, assume that the best
window size is th...