You are on page 1of 8

INDUSTRIAL APPLICATIONS

Alternative Mathematical
Approaches to Shade Sorting

J. R. Aspland,1 K. D. Balasaygun,2
J. P. Jarvis,3* T. H. Whitaker3
1
School of Textiles, Fiber and Polymer Science, Clemson University, Clemson, SC

2
AT&T Laboratories, 200 Laurel Ave., Middletown, NJ 07748

3
Department of Mathematical Sciences, Clemson University, Clemson, SC

Received 6 November 1998; accepted 8 November 1999

Abstract: Shade sorting is the process of assigning samples garment parts are cut from different rolls and sewn together
of the same nominal color into groups having no significant without considering this variation, the end product may be
color variation. Use of modern spectrophotometers and unsatisfactory.
color measurement technology make it possible to obtain One method used to avoid this problem is a process
precise color differences between samples. When these known as shade sorting. This operation is often performed
color differences are viewed as distances between points, by trained color sorters, who visually group the rolls of
the shade sorting problem is seen to be equivalent to the fabric to assure that those rolls cut together are similar
clustering problem in the mathematical literature. Several enough in shade to be sewn into adjacent panels of a single
mathematical techniques for clustering— complete linkage garment.1 However, advances in technology have produced
clustering, vertex labeling, and set covering—are explained color measurement instruments with sufficient accuracy and
and compared for their efficiency when applied to shade reliability to be used to measure color for numerical shade
sorting. A particular implementation of complete linkage sorting. Numerical shade sorting refers to quantitative, ob-
clustering called Clemson Color Clustering (CCC) is found jective methods for grouping rolls of fabric of similar color.
to perform well as compared to the other reviewed It avoids the subjective aspects of visual shade sorting.
methods. © 2000 John Wiley & Sons, Inc. Col Res Appl, 25, 368 –375, Several different mathematical techniques can be used for
2000 numerical shade sorting. These techniques fall under the
general category of clustering. Three common approaches
Key words: shade sorting; clustering; color measurement; to the clustering problem— complete linkage clustering,
Clemson Color Clustering; CCC vertex labeling, and set covering—are explained below.
These methods are then compared for their usefulness in
INTRODUCTION numerical shade sorting for the apparel industry.

A common quality problem in cut and sew operations is


variation in color from one roll of fabric to another. If NUMERICAL SHADE SORTING METHODS

In the apparel business, the goal of numerical shade sorting


is to divide a set of samples into a minimum number of
* Correspondence to: Prof. James P. Jarvis, Department of Mathematical
Sciences, Martin Hall, Clemson University, Clemson, SC 29634 (e-mail: groups (cutting lots) such that the color range of each group
jpjrv@clemson.edu) is no larger than a specified maximum color difference, the
© 2000 John Wiley & Sons, Inc. cut tolerance.2 This tolerance is the maximum permissible

368 COLOR research and application


color difference between adjacent panels in a garment. The mathematical problem known as the clustering problem.8
same principles also apply to any industry producing a The formal statement of the clustering problem is: Given a
multiplicity of colored products that might end up adjacent finite set of points X, an integer distance function defined on
to one another—for example, rolls of wallpaper. pairs of points from X, a positive integer k, and a positive
integer D, is there a partition of X into k sets, each of which
has diameter of at most D. (The diameter of a set is defined
Measuring Color Differences
to be the maximum distance between two elements of that
Any color space and color difference equation can be set.)
used to calculate the color difference between colored sam- Intuitively, clustering attempts to group objects so that
ples (data points). The mathematical procedures described items in the same group tend to be similar to each other.
below require only that the color difference between pairs of Specifically, the objective is to partition a given set of n
points be known. The procedures operate independently of entities into homogeneous and well-separated subsets called
how the difference is calculated. Of course, to ensure sat- clusters.
isfactory results, the color difference equations used should The clustering problem is contained in a class of prob-
give color differences that correspond well with human lems that are called NP-complete.8 No known efficient
visual perception. solution procedures exist for such problems. Generally,
One method for calculating color difference that has heuristic algorithms must be employed to obtain good,
performed well in the past is the CMC(ᐉ:c) color difference although not necessarily optimal, solutions, since all known
equation. 3 To correct for the nonuniformity of color space, methods for solving such problems require exponentially
CMC determines a set of scaling factors for lightness, increasing time as problem size increases.
chroma, and hue from a given standard color so that an The numerical shade sorting methods discussed here at-
equal distance in any direction for that standard is equally tempt to solve the clustering problem. All are equivalent and
acceptable for color. These scaling factors determine an each comes from a different mathematical discipline. Com-
elliptical region around the standard, whose surface consists plete linkage clustering is a statistical method; vertex label-
of points in color space that are equally acceptable as the ing is an approach from graph theory; and set covering is an
standard. The points contained in this ellipsoid of color operations research technique. Each of these methods is a
space correspond to goods that are close in color to the general heuristic that is adapted here to the shade sorting
standard. problem.
Again, using the apparel cutting and sewing industry as
an example, a single cut tolerance is generally not accept-
able for all applications of color sorting due to differences in Clemson Color Clustering
the material, customer acceptance levels, visual perception,
etc. Setting tolerances is often achieved by trial and error. In 1985, Aspland et al.6 introduced a clustering procedure
Record keeping can be used to set provisional tolerances, or called Clemson Color Clustering (CCC), which uses a tech-
the CMC(ᐉ:c) color difference formula can be used to set nique known as hierarchical complete linkage clustering.9
starting values. Once the shade sorting system is put to use In this method and those discussed below, distance and
in a practical application, tolerances can be adjusted as color difference are used interchangeably. Complete linkage
needed.4 refers to the property that the distance between two clusters
The shade sorting results presented in the following ma- is defined to be the maximum color difference between any
terial employ CIELCH 1976 colorimetric values5 and use two rolls, one from each cluster. The diameter of a cluster is
the CMC (2:1) color difference formula in the calculation of defined to be the maximum color difference over all pairs of
the color differences. For simplicity of exposition in the rolls in that cluster.
remainder of this article, lightness, chroma, and hue are To cluster a given set O ⫽ {O1, O2, . . . , On} of n objects
scaled by their associated color-scaling factors so that color using this technique, O is initially partitioned into n single-
ellipsoids can be pictured as spheres. roll clusters, denoted Pn. Each iteration involves merging
two clusters that produce a new cluster of minimum diam-
eter. This process continues until all rolls have been merged
Numerical Shade Sorting
into a single cluster or until no more clusters can be com-
Until 1985, numerical shade sorting was most often based bined without exceeding the cut tolerance. In this way, a
on a fixed-grid method for subdividing a volume of color hierarchy {Pn, Pn⫺1, . . . , Pk} of partitions of O into n, n ⫺
space into smaller volumes. This was called the 555 meth- 1, . . . , k clusters is produced such that for each pair of
od.1 In fixed-grid methods, the smaller volumes that contain partitions Pi⫹1 and Pi, each cluster of Pi⫹1 is a cluster of Pi,
data points represent shade groups and are fixed shapes that except that a pair of clusters of Pi⫹1 is merged to a new
partition the acceptable volume of color space. Such a single cluster of Pi. When two clusters are merged, they are
division of color space depends on the coordinates of the permanently joined.
standard, but does not take the positions of sample data into Figure 1 illustrates how this method successively groups
account.6,7 the closest clusters to obtain larger clusters. Four clusters
Numerical shade sorting can also be formulated as a are successively combined to ultimately produce a single

Volume 25, Number 5, October 2000 369


In stage 1, a graph is constructed. The graph consists of
vertices that represent samples to be shade sorted, with an
edge between each pair of vertices whose distance in color
space exceeds the specified color tolerance. Each vertex is
assigned a label such that no two adjacent vertices receive
the same label. This labeling partitions the points into a
minimal number of groups, each with diameter less than the
cut tolerance.
In stage 2, new edges are added to the graph in order of
decreasing color difference, starting at the cut tolerance.
After the addition of each edge, an attempt is made to

FIG. 1. Example of groupings produced by complete link-


age method for 100 samples of a green knit fabric. Top
number indicates number of sample points and bottom
number indicates maximum color difference within each
cluster.

cluster. The final clusters produced by the complete linkage


algorithm determine the shade sorting groups.
The major criticism of complete linkage cluster analysis
is that clusters tend to recede upon growth. As the size of a
cluster increases, the effective color difference between that
cluster and any entity not contained in that cluster also
increases. Hence, the probability of a cluster obtaining a
new member becomes smaller as the size of the cluster
increases.10 This property is an asset in terms of shade
sorting, because it tends to produce groups uniform in size
as opposed to a few large clusters and many small clusters.

Vertex Labeling Approach


A graph consists of a vertex set (points or objects) and an
edge set (pairs of objects). Each edge is thought of as
joining a pair of vertices. Two vertices are called adjacent if
and only if they are joined by an edge.8 Vertex labeling
attempts to label each vertex of a graph with a minimum
number of labels so that no two adjacent vertices have the
same label.
Vertex labeling, with an appropriately defined graph, is
equivalent to the clustering problem defined previously. The
rolls to be sorted are represented by the vertices of the
graph. An edge is placed between vertices for which the
color difference is greater than the cut tolerance. Labels are
assigned to all vertices such that all adjacent vertices have
different labels. In this way, labeling the vertices of the
graph serves to partition the rolls of fabric into shade
groups, each with diameter less than or equal to the cut FIG. 2. Vertex labeling algorithm applied to a set O of four
tolerance. Vertices with a common label correspond to a entities. (a) Color difference between each pair of entities. (b)
single shade group. In stage 1: the partial graph G2.5 is labeled with three labels.
Hansen and Delattre11 have shown that this problem is (c) In stage 2: add the next largest edge to G2.5 and relabel.
(d) Create G2.0. (e) Add the next largest edge to produce
reducible to the problem of optimally labeling a sequence of graph G1.0 that cannot be labeled using only three labels.
graphs. The 2-step procedure that follows is a modified G2.0 yields a minimal diameter partition of the set O into
version of their algorithm. three groups.

370 COLOR research and application


relabel the new graph using the same number of labels.
Vertices must be relabeled, if an edge joins two vertices in
the same group. If the graph can be relabeled, the maximum
diameter of the resulting groups is reduced. This process
continues until the graph cannot be labeled without increas-
ing the number of labels. This stage tends to produce groups
uniform in size as opposed to one large group and several
small groups.
The vertex labeling algorithm is illustrated in Fig. 2. A set
O of 4 entities is to be clustered, allowing a maximum
diameter (cut tolerance) of 2.5 within a group. The distance
matrix D shows the color difference between each pair of
entities [Fig. 2(a)]. In stage 1, the associated graph G2.5 is
formed. Applying the vertex labeling algorithm to the graph
produces the labels indicated [Fig. 2(b)]. Three labels are
required, yielding a maximum diameter of 2.0. Hence, the
entities are sorted into three groups. The second stage of the
algorithm performs the diameter reduction. First, edge (c, a)
with a weight of 2.0 is added to the graph, producing the
graph G2.0 [Fig. 2(c)]. Since vertices c and a both have label
2, relabeling is necessary. Vertex c cannot be relabeled,
because it is already adjacent to vertices of all three labels.
Vertex a, however, is adjacent to only two of the three
labels and is reassigned the label 3 [Fig. 2(d)]. The new
partition has a maximum diameter of 1.0. Adding the next
largest edge (d, a) to the graph G2.0 produces the graph G1.0,
which cannot be labeled with only three labels [Fig. 2(e)].
Therefore, the final labeling of the graph G2.0 yields a
minimal diameter partition of the 4 entities into 3 groups.

Set Covering Approach

The shade sorting problem can also be modeled as a set FIG. 3. Set covering algorithm applied to four sample
covering problem.12 For shade sorting, the goal is to posi- points. (a) The coverage matrix for the given data points. (b)
Reduction 1 applied to the coverage matrix. (c) Reduction 2
tion a minimum number of spheres in color space such that applied to the coverage matrix. (d) Solution: centers at
these spheres collectively contain (cover) all sample points. points w and z.
Each sphere is centered at a sample point. Hence, this
method determines a minimal number of sample points to
serve as centers of spheres, so that each sample point is this method is computationally expensive.13 An effective
contained in at least one sphere. In set covering terms, the heuristic procedure is given below.
rolls of fabric correspond to the entities to be covered. The The algorithm that follows uses a series of logical reduc-
radius of each covering sphere is half the cut tolerance, so tions12 to produce successively smaller problems (reducing
that the maximum color difference between any two points the size of A) until a final solution is obtained. Each reduc-
covered by a sphere does not exceed the cut tolerance. The tion is performed until it is no longer possible to use that
sample points contained in a sphere define a shade group. reduction. The entire set of reductions is repeated until none
First, the color difference of each pair of sample points is can be applied.
calculated and the coverage matrix A ⫽ (aij) is formed. The
rows of A correspond to possible centers of spheres, and the Reductions:
columns correspond to sample points to be covered by the 1. If there is a sample point covered by only one potential
spheres. Initially, each sample point is considered a poten- center, then a sphere must be centered at this point. This
tial center for a covering sphere. A sample point is covered row is eliminated, and each column corresponding to a
by a potential center, if the color difference between this point covered by this sphere is eliminated.
point and the point corresponding to the potential center 2. If sample i is covered whenever sample j is covered,
does not exceed half the cut tolerance. Hence, the (i, j) entry then point i can be ignored. Eliminate column i.
in the coverage matrix is a 1, if point j is covered by 3. If a potential center j covers at least the same points as
potential cover i. Otherwise, entry (i, j) is a 0. This problem a potential center i, then potential center j is preferred.
can be solved using general integer linear programming, but Eliminate row i.

Volume 25, Number 5, October 2000 371


FIG. 4. Number of shade groups produced by complete FIG. 5. Number of shade groups produced by complete
linkage clustering (CCC), vertex labeling, and set covering linkage clustering (CCC), vertex labeling, and set covering
when applied to 100 rolls of a green knit fabric. when applied to 149 rolls of a beige fabric.

If the matrix is completely reduced (to a 0 ⫻ 0 matrix) by groups produced by each of the three methods over a range
these steps, the solution obtained is optimal. Otherwise, of color tolerances.
another procedure must be used so that the reductions can Some general patterns are evident from these inventories.
be applied again. One reasonable procedure is to choose the With few exceptions, the vertex labeling approach produces
center that covers the largest number of remaining sample the smallest number of groups for a given color tolerance,
points and designate it as a cover, then eliminate the corre- while set covering performs consistently worse than com-
sponding row and all columns corresponding to points cov- plete linkage clustering and vertex labeling. Hence, in terms
ered by this sphere. This heuristic allows the matrix to be of number of groups, vertex labeling appears to be some-
fully reduced, although the solution so obtained is not what superior for shade sorting.
necessarily optimal. The set covering algorithm is illus-
trated in Fig. 3.
The exact number and position of the covering spheres Uniformity in Size and Color Range
depend on the particular sample data. Although the diameter A goal of shade sorting is to produce groups that are
of each covering sphere is fixed and equal to the cut toler- uniform in size and color consistency in order to utilize all
ance, the actual diameter of the shade group produced by a goods and minimize waste in a cutting spread. To be used
covering sphere may be considerably less than the cut for processing a cut, a shade group must be large enough to
tolerance. For instance, in the previous example, the diam- meet the yardage requirement of the desired cut. Hence, the
eter of the group containing only point a is zero. relative size of the groups produced by a shade sorting
method is important. In addition, it is desirable to sort the
samples into groups of minimum diameter. These two char-
EXPERIMENTAL COMPARISON OF ALGORITHMS acteristics tend to work together. As the size of a shade
Datasets from actual textile inventories are used to compare group increases, its diameter also tends to increase. Ideally,
shade sorting by complete linkage clustering, vertex label-
ing, and set covering. Using the color inventories, the shade
groups produced by each of the three methods are compared
for their number, anomalies, and uniformity in size and
color range. The methods themselves are compared for
computational complexity of the algorithms, execution
time, and ability to combine production lots for color in-
ventory management.

Number of Groups
One basis for experimentally comparing the different
shade sorting methods is the number of groups produced for
a given color tolerance. The three approaches— complete
linkage clustering, vertex labeling, and set covering— have FIG. 6. Number of shade groups produced by complete
been used to analyze three color inventories. Results are linkage clustering (CCC), vertex labeling, and set covering
given in Figs. 4 – 6. These figures show the number of when applied to 88 rolls of a gray fabric.

372 COLOR research and application


TABLE I. Complete linkage clustering (CCC), vertex labeling, and set covering applied to 149 rolls of a beige
fabric using a cut tolerance of 0.700. Diameter and number of samples are given for each group.
CCC Vertex labeling Set covering
Diameter Size Diameter Size Diameter Size

0.564 35 0.677 43 0.670 43


0.537 26 0.673 30 0.677 41
0.497 16 0.677 23 0.646 40
0.456 16 0.685 11 0.636 37
0.660 15 0.673 8 0.653 36
0.480 9 0.649 8 0.620 15
0.569 7 0.676 7 0.630 12
0.481 7 0.591 7 0.638 11
0.353 7 0.679 6 0.558 7
0.555 4 0.212 6 0.591 7
0.465 4 0.278 6
0.557 2 0.533 5
0.000 1 0.296 2
0.000 1
0.000 1

if n points were partitioned into m groups, each group would Note that some additional processing of shade groups
contain n/m points. might be advantageous after the initial clustering is per-
A comparison of the three shade sorting methods— com- formed. For example, consider the groups indicated in Fig.
plete linkage clustering, vertex labeling, and set cover- 7. Suppose that eight rolls of fabric are needed to process a
ing—is shown in Table I. A cut tolerance of 0.70 was used cut. Group A does not contain enough rolls, while Group B
to shade sort an inventory consisting of 149 rolls of a beige contains more than eight. By examining the data, we see
fabric. This table indicates the diameter and size (number of that the cut can be processed by taking the rolls from Group
rolls) of each group produced. These results are typical of A along with three rolls from Group B. As this example
those obtained with other fabrics and cut tolerances. In illustrates, relatively small groups are not necessarily a
Table I, only a slight difference can be seen in the number problem. Although it is easier to use groups containing a
of groups produced by the methods. Set covering produced large number of rolls, small groups can be used efficiently,
two singleton groups, as opposed to only one singleton if the data are examined after shade sorting is completed.
group in complete linkage clustering and none in vertex
labeling. All three methods tend to distribute the samples
Complexity of the Algorithms
well resulting in good colorimetric uniformity within the
shade sorting groups. Each of the three shade sorting algorithms— complete
In general, the hierarchical complete linkage algorithm linkage clustering, vertex labeling, and set covering—re-
used in Clemson Color Clustering may group very similar quires space in the order of n2 to store the color difference
entities in the same cluster only at a late stage in the data for a problem of size n (n rolls). The worst case
algorithm. Occasionally, this can lead to inhomogeneity in theoretical complexity of sorting n entities is of the order n3
some clusters. for complete linkage clustering and of the order of n4 for
The vertex labeling method also tends to produce groups both vertex labeling and set covering.7
that are relatively uniform in size and color range. In large
part, this is a result of stage 2 of the vertex labeling algo-
rithm, which attempts to reduce the maximum diameter of
the groups. By distributing the points more evenly through-
out the groups, the maximum color variation within a shade
group is reduced without increasing the number of groups.
Unlike complete linkage clustering and vertex labeling,
the sole objective of the set covering method is the mini-
mization of the number of groups produced. The set cover-
ing approach ensures that the diameters of the covers do not
exceed the cut tolerance, but the algorithm is not designed
to minimize these diameters. The set covering method does
not uniquely assign points to groups. Each sample is con-
tained in at least one group, possibly more. As a result, the
FIG. 7. Shade groups A and B. The circles represent the
diameters and sizes of the individual groups tend to be maximum diameter of a cluster allowed by the cut tolerance.
larger than those produced by the other methods, which The rolls in the intersection of the circles could actually
partition the points into mutually exclusive groups. reside in either group.

Volume 25, Number 5, October 2000 373


TABLE II. Computation time (s) required for shade groups that are fixed relative to the standard. However, the
sorting on a personal computer. Three datasets were center of each shade group can be calculated to provide an
used: a gray corduroy inventory containing 88 rolls, a indication of the locations in color space of the shade groups
green knit inventory containing 100 rolls, and a beige relative to the standard.
inventory containing 149 rolls t (The same data are In many situations, the color relative to the standard is not
used in Fig. 4 – 6.) Each inventory is sorted at a cut a major concern. One of the most important uses of shade
tolerance of 1.0 using the complete linkage clustering,
sorting is to incorporate unused fabric from one lot into
vertex labeling, and set covering methods.
another lot of the same production standard. In complete
Time required (s) linkage clustering, colorimetric data from a new lot may be
Sorting method n ⫽ 88 n ⫽ 100 n ⫽ 149 added to the data for unused fabric from previous lots and
simply reclustered. In the vertex labeling method, a vertex
CCC 4 5 11 may be added to the graph for each new sample point and an
Vertex labeling 9 11 22
Set covering 108 113 259 edge added between this vertex and each existing vertex
whose corresponding color difference exceeds the cut tol-
erance. Relabeling the resulting graph yields the new shade
groups. In set covering, the new colorimetric data points are
Each algorithm has been implemented on a personal added to the existing data. If the new data points lie entirely
computer to compare their execution times empirically. To within existing covers, resorting is unnecessary. Otherwise,
estimate the efficiency of each algorithm, the same three the set covering method is applied to the combined data and
sets of textile data as described above were shade sorted new covers are formed.
with the computer programs. The computation times (s) for Resorting does not necessarily produce the same shade
each of the inventories over a range of cut tolerances are groups as those initially obtained. However, as goods are
given in Table II. In computation time, all these examples produced and consumed, resorting combines the inventory
perform significantly better than the predicted worst-case into new groups that can be used more effectively in pro-
complexity. The times do not appear to grow in the order of duction. The main problem is not shade sorting, but main-
n3 or n4, as expected. Rather, each method tends to require taining the inventory in such a way that each roll can be
a time in the order of n2. Overall, the empirical ratio in run located and selected regardless of its shade group. The
times between complete linkage clustering and vertex label- benefit of resorting is the avoidance of a fragmented inven-
ing is about 1:2, with the time required for set covering tory—many remnant groups are left with only a small
being considerably larger than the other two methods. Note number of rolls after initial use of the group. In fact,
that the ratio between times is probably more significant judicious choice of rolls for cutting can actually compact an
than the absolute times, which depend on the characteristics inventory into fewer shade groups. Complete linkage clus-
of the particular personal computer being used. tering, as implemented in CCC software, with color inven-
tory management, implements such techniques.14
Clustering Anomalies
The vertex labeling and set covering methods can actu- CONCLUSIONS
ally produce a larger number of groups for a dataset as the
maximum allowable color range is increased. This is be- Overall, each of these mathematical methods— complete
cause the vertex labeling algorithm is applied to a different linkage clustering, vertex labeling, and set covering—pro-
graph and the set covering algorithm is applied to a different duces a reasonably small number of groups, good color
coverage matrix when the cut tolerance is increased. Such a uniformity within the groups, and good uniformity in the
situation cannot occur in complete linkage clustering, be- number of samples contained in each group. Complete
cause the clustering procedure is hierarchical. In complete linkage clustering and vertex labeling are superior to set
linkage clustering, pairs of clusters are merged until the cut covering in the number of groups produced. The three
tolerance is exceeded. Increasing the cut tolerance corre- approaches are similar in the uniformity of groups.
sponds to simply terminating the algorithm at a later stage. Although complete linkage clustering was slightly infe-
The same intermediate clusters are formed. rior to the vertex labeling algorithm in these examples, the
small gain in number, color uniformity, and size of groups
is offset by the more lengthy execution times associated
Combining Production Lots
with vertex labeling. In addition, both vertex labeling and
The shade groups produced by these methods are based set covering can generate a larger number of groups for the
on the color difference between samples, not the actual same dataset when the cut tolerance is increased, while
location of samples in color space. Consequently, the shade complete linkage clustering does not display this anomaly.
groups are not fixed from one production lot to another. Overall, complete linkage clustering and vertex labeling are
Samples grouped together do not necessarily remain to- efficient mathematical approaches to numerical shade sort-
gether after the addition or deletion of inventory and the ing and produce shade groups with a number of desirable
subsequent resorting. Similarly, these methods do not form properties.

374 COLOR research and application


1. Simon FT. Shade sorting by the 555 system. Am Dye Rep 1984;73: 8. Garey MR, Johnson DS. Computers and intractability. San Francisco:
17–26. WH Freeman; 1979.
2. Li YSW, Yuen CWM, Yeung KW, Sin KM. Instrumental shade 9. Kaufman LK, Rousseeuw PJ. Finding groups in data: An introduction
sorting in the past three decades. JSDC 1998;114:203–209. to cluster analysis. New York: Wiley; 1990.
3. Clarke FJJ, McDonald R, Rigg B. Modification to the JPC79 colour- 10. Blashfield RK. Mixture model tests of cluster analysis: Accuracy of
difference formula. JSDC 1984;100:128 –132. four agglomerative hierarchical methods. Pysch Bull 1976;83:377–388.
4. Aspland JR, Jarvis CW, Jarvis JP. A review and assessment of nu- 11. Hansen P, Delattre M. Complete-link cluster analysis by graph color-
merical shade sorting methods. JSDC 1990;106:9 –14. ing. J Am Stat Assoc 1978;73:397– 403.
5. Billmeyer F, Saltzman M. Principles of color technology. New York: 12. Balinski M, Quandt R. On an integer program for a delivery problem.
Wiley; 1981. Opns Res 1964;12:300 –304.
6. Aspland JR, Jarvis CW, Jarvis JP. An improved method for numerical 13. Garfinkel RS, Nemhauser GL. Integer programming. New York:
shade sorting. Text Chem Colorist 1987;19:21–25. Wiley; 1972.
7. Dunlap KL. Comparison of methods for numerical shade 14. Teel DR, Aspland JR, Jarvis JP, Dunlap KL. Improved methods for
sorting. Clemson Univ: Dept Math Sci Doctoral Dissertation; colour inventory management in the apparel industry. Int J Clothing
1992. Sci Tech 1992;4:66 –70.

Volume 25, Number 5, October 2000 375

You might also like