RETRO commands


PROGRAM NAME

vwmdd - Weighted Minimum Distance Detector (K1)

DESCRIPTION

vwmdd is an Weighted minimum distance detector or Classifier. The main idea behind this detector is that it will distinguish a single class form the rest of the data. However, it will also work to distinguish multiple classes from each other.

Theory

To classify or detect spatial patterns in an image, we must have a template to compare against. Creating a template can be thought of as training on representative (ideal) data. This can be done by using a clustering algorithm on the representative data. Clustering routines such as vkmeans, vquant, isodata, etc can be used to create a template. The idea here is to over cluster the representative data, and quantize that into k classes. Pseudo color in editimage, and kasc2val can be used to create a class image. When the clustering is performed, two items need to be computed; (1) the cluster center values, and (2) the variances for each axis in the cluster. The variances can be obtained by looking at the diagonal of the covariance matrix.

Explained below is a method of creating a the class assignments for each cluster center. The class assignments may, for example, quantize the space from 12 clusters to 2 classes. Example


Cluster Center Class 1 --------------------- 2 2 --------------------- 2 3 --------------------- 1 4 --------------------- 2 . . . 12 --------------------- 1

This routine expects an image that contains the diagonal terms of the covariance matrix, called the variance image, and an image that contains the cluster center values and class assignment for each cluster center. And of course the image to be classified. This detector is to approximate the following equation:


(X - Mi)~ Inv(C) (X - Mi) (1)

where Mi is the mean, inv(C) is the inverse of the Covariance matrix, and X is the data points, and ~ denotes transpose. This can be written as

(d1 d2 ... dn)~ - Q1 - (d1 d2 ... dn) (2) | Q2 | | . 0 | | . | | 0 . | | Qn | - -

Which equals

sq(d1) sq(d2) sq(dn) ------ + ------ + ..... + ------ (3) Q1 Q2 Qn Notation: sq => square Qi => the ith variance di => the ith (X - Mi) value the matrix is actually the inverse of C

Since the inverse of the Covariance matrix can not be easily determined, we only consider the diagonal terms or simply the variance terms.

There are two methods for detecting classes in the input image. Each data point is put through a test based on the ideas above. The detector works in two modes as follows:

(1) uses the summed method -s option set to yes. This is an approximation to method (2).


sq( || X - Mi || ) < Vi = ------------------ > 1 K x sq(Si)

where: sq( || X - Mi || ) = sq(di) = Euclidean Distance Squared sq(Si) = trace(C) the trace of C is the sum of the diagonal terms (variances). K is a Constant

(2) non summed method (default).


sq(d1) sq(d2) sq(dn) < Vi = ------ + ------ + ..... + ------ > 1 Q1 x K Q2 x K Qn x K

where: sq(dj) = sq( || xj - Mi || ) = Euclidean Distance Squared of the jth element of the vector X. Qj = the jth variance element of the variance vector for the ith cluster center. K = Constant

In both cases the constant K is used. This is used to adjust the variance value(s). If K is large, Qi can increase such that Vi is always less than 1. There are no specific rules as to the optimal value for K.

The data point X is put through the test for each cluster, the data point X is assigned to a Class based on the following criteria:

1. choose the smallest Vi value, where Vi is the result for cluster i.

2. If Vi is greater than 1, assign X to the NULL class (NULL class is always 0). Otherwise assign Vi to class that corresponds to the ith cluster.

This routine accepts three images as the input. The image that corresponds to the -i1 argument is the image that needs to be classified. This image can be have as many data bands as necessary. The number of data bands depicts the dimensionality of the vectors. Each pixel in the image is a vector with dimensionality of the number of data bands.

The second input image corresponds to the -i2 argument and is the prototype image. This image contains two pieces of information. (1) the cluster center values for each cluster center, and (2) the class assignments for each cluster center. The last data band of this image contains the class assignments for each cluster center value. This image must contain the same number of data bands as the other input image plus an extra data band that represents the class mapping. This image would most likely have been created by vkmeans or some other routine that will give cluster centers. This image contains vectors that correspond to the prototype of each class.

As stated above the center image's last data band is a class data band. The class data band simply maps each cluster center vector to a final class.

At this point most class images must be created manually. A class image can be created by using pseudo color in editimage on the resulting clustered image from the vkmeans routine. Vcustom can then be used to create the class image. Finally vinset can be used to add the class image data to the center image. Normally one would over cluster using vkmeans then reduce the space down. So vkmeans might produce 100 clusters but, really only 3 classes are desired. Pseudo color in editimage and kasc2val can be used to reduce the space down and create a class image. The class data might map cluster center vectors 1-50 to class 1, vectors 51-60 to class 2 and vectors 61-100 to class 3. When the vwmdd routine is run, the result would be a classified image of three classes.

The third input image corresponds to the -i3 argument. This image contains the variance values for each cluster center. This image contains the same number of data bands as the input image (-i1 image), and contains the same number of vectors as the center image (-i2 image). Each element in the variance vector is the variance value along an axis in the n-dimensional space. The variance values can be found by computing the diagonal of the covariance matrix for each cluster. The cluster center values would be used as the mean values. Note, vkmeans will compute this image for you.

The scale factor (-k) must be chosen by trial and error. It has been found that if the summed method is being used a small (1.0 - 3.0) value for k is sufficient. If the average method is not being used a larger k (8.0 - 10.0) value is sufficient. Note, this may change based on the data, so these ranges are only a starting point.

The border option (-b) allows the user to specify a border width, in pixels, encompassing the image. The border region is skipped by vwmdd when classification is performed. This useful if neighborhood operators have been used previously and have not updated edge pixels.

All input images must be of data storage type FLOAT.

This routine was written with the help of and ideas from Dr. Don Hush, University of New Mexico, Dept. of EECE.

REQUIRED ARGUMENTS

-i1
type: infile
desc: Input Image
-i2
type: infile
desc: Input Cluster Center/Class Image
-i3
type: infile
desc: Input Variance Image
-o
type: outfile
desc: Output specifying which vector belongs to which cluster

OPTIONAL ARGUMENTS

-k
type: float
desc: Scaling factor
default: 1
bounds: value > 0.0
-b
type: integer
desc: Border width in pixels
default: 0
bounds: 0 < [-b] < 100
-s
type: boolean
desc: If true, Use Summed Method
default: false

EXAMPLES

vwmdd -i1 aerial_image.xv -i2 centers_class -i3 variance -b 5 -k 8.0 -o test

aerial_image is a 5 band image; thus, each vector has dimensionality of 5, centers_class image is a 6 band image, 5 bands that contain the cluster center values and 1 class band. The border that vwmdd will ignore is 5 pixels wide. The output image will be a single band image called test. The scalar to multiply the denominator by is 8.0. By default the non-summing method is used.

SEE ALSO

vkmeans(1)

RESTRICTIONS

All input images must be of data storage type FLOAT.

REFERENCES

COPYRIGHT

Copyright (C) 1993 - 1997, Khoral Research, Inc. ("KRI") All rights reserved.