Reprinted by permission from IEEE TRANSACTIONS ON SYSTEMS, MAN AND CYBERNETICS

Vol. SMC-3, No.6, November 1973, pp. 610-621

Copyright 1973, by the Institute of Electrical and Electronics Engineers, Inc.
PRINTED IN THE U.S.A.

Textural Features for Image Classification

ROBERT M. HARALICK, K. SHANMUGAM, AND ITS'HAK DINSTEIN

Abstract-Texture is one of the important characteristics used in array. If L, = {1,2, · · · ,N,} and L, = {1,2, · · · ,N,} are the
identifying objects or regions of interest in an image, whether the image X and Y spatial domains, then L, x L, is the set of resolu-
be a photomicrograph, an aerial photograph, or a sateDite image. This
paper describes some easily computable textural features based on gray-
tion cells and the digital image /is a function which assigns
tone spatial depeodancies, and illustrates their application in category- some gray-tone value GE {1,2, · · · ,N,} to each and every
identification tasks of three different kinds of image data : photo- resolution cell; /: Lx x L, --+ G. Various two.dimensional
micrographs of five kinds of sandstones, 1:20 000 panchromatic aerial analyses are performed on I to achieve specific image-
photographs of eight land-use categories, and Earth Resources Tech- processing tasks such as coding, restoration, enhancement,
nology Satellite (ERTS) multispecial imagery containing seven land-use
categories. We use two kinds of decision rules: one for which the decision
and classification. In recent years a tremendous amount of
regions are convex polyhedra (a piecewise linear decision rule), and one computer processing of photographs has occurred, with
for which the decision regions are rectangular paraUelpipeds (a min-max facilities having been developed to process anything from
decision rule). In each experiment the data set was divided into two aerial photographs to photomicrographs [!], [2].
parts, a training set and a test set. Test set identification accuracy is
In this paper we are concerned with the task of developing
89 percent for the pbotomicrograpbs, 82 percent for the aerial photo-
graphic imagery, and 83 percent for the satellite imagery. These results
a set of features for classifying or categorizing pictorial
indicate that the easily computable textural features probably have a data. The classification of pictorial data can be done on a
general applicability for a wide variety of image-classification applica• resolution cell basis (such as in identifying the crop category
tioos. of a resolution cell on satellite imagery) or on a block of
contiguous resolution cells (such as in identifying the crop
I. INTRODUCTION
category of an entire agricultnral field extending over a

W ITH THE ADVENT of high-speed general-purpose

digital computers it is becoming possible to perform
mathematical or algorithmic processes on pictorial data
large number of resolution cells). The most difficult step in
categorizing pictorial information from a large block of
resolution cells is that of defining a set of meaningful
from images of photographic quality. In most of these features to describe the pictorial information from the block
processes, the pictorial information is represented as a of resolution cells. Once these features are defined, image
function of two variables (x,y). The image in its digital blocks can be categorized using any one of a multitude of
form is usually stored in the computer as a two-dimensional pattern-recognition techniques.
In a search for meaningful features for describing pic-
Manuscript received April 25, 1973; revised June J2, 1973. This
torial information, it is only natural to look toward the
work was supported by the NASA Goddard Space Flight Center under types of features which human beings use in interpreting
Contract NASS-21822. and the Geographic Sciences Laboratory of pictorial information. Spectral, textural, and contextual
the U.S. Army Engineer Topographic Laboratories under Contract
DAAK02-70-C-0388. features are three fundamental pattern elements used in
R. M. Haralick and I. Dinstein are with the School of Electrical human interpretation of color photographs. Spectral fea-
Engineering and the Center for Research, Inc., University of Kansas,
Lawrence, Kans. 66044. tures describe the average tonal variations in various bands
K. Shanmugam was with the Center for Research, Inc., University of of the visible and/or infrared portion of an electromagnetic
Kansas, Lawrence, Kans. He is now with the Department of Electrical
Engineering, Wichita State University, Wichita, Kans. 67218. spectrum, whereas textural features contain information
HARALICK et al.: TEXTURAL FEATURES FOR IMAGE CLASSIFICATION
611

about the spatial distribution of tonal variations within a Rosenfeld and Troy described a procedure for obtaining a
band. Contextual features contain information derived from measure of the texture "coarseness" of images. Their pro-
blocks of pictorial data surrounding the area being analyzed. cedures were based on the differences between the gray-tone
When small image areas from black and white photographs values of adjacent image elements and on the autocorrela-
are independently processed by a machine, then texture and tion of the image gray-tone values. In [8] Rosenfeld and
tone are most important. Thurston gave a procedure for detecting boundaries separat-
The concept of tone is based on the varying shades of ing regions which differ in texture coarseness. In [9] Troy
gray of resolution cells in a photographic image, while et al. described gray-level manipulation procedures which
texture is concerned with the spatial (statistical) distribution can be used for preprocessing of pictures before applying
of gray tones. Texture and tone are not independent con- the algorithms given in [7] and [8]. Procedures for detecting
cepts; rather, they bear an inextricable relationship to one textural properties such as lines and dots have also been
another very much like the relationship between a particle suggested by other investigators (see, for example, [10]-
and a wave. Context, texture, and tone are always present in (12]). Before applying these procedures to pictures other
the image, although at times one property can dominate the than the ones processed by the authors of the respective
other. papers, the investigator has to make a choice as to the
Texture can be evaluated as being fine, coarse, or smooth; method appropriate to the picture in question as well as the
rippled, moiled, irregular, or lineated. For example, in the selection of parameters for the particular method.
humid tropics, fine texture on radar imagery can be in- We are presenting in this paper a general procedure for
dicative of nonresistant fine-grain sedimentary rocks and extracting textural properties of blocks of image data. These
unconsolidated sediments, while a coarse texture can be features are calculated in the spatial domain, and the
indicative of coarser grained sedimentary rocks. A massive statistical nature of texture is taken into account in our
texture with high-contrast components may be indicative of procedure, which is based on the assumption that the tex-
igneous rocks. A hummocky texture can be indicative of ture information in an image / is contained in the overall
eroded igneous rocks. or '"average" spatial relationship which the gray tones in
Texture is an innate property of virtually all surfaces- the image have to one another. We compute a set of gray-
the grain of wood, the weave of a fabric, the pattern of tone spatial-dependence probability-distribution matrices
crops in a field, etc. It contains important information for a given image block and suggest a set of 14 textural
about the structural arrangement of surfaces and their features which can be extracted from each of these matrices.
relationship to the surrounding environment. Although it These features contain information about such image tex-
is quite easy for human observers to recognize and describe tural characteristics as homogeneity, gray-tone linear de-
in empirical terms, texture has been extremely refractory to pendencies (linear structure), contrast, number and nature
precise definition and to analysis by digital computers. of boundaries present, and the complexity of the image. It
Since the textural properties of images -appear to carry is important to note that the number of operations required
useful information for discrimination purposes, it is im- to compute any one of these features is proportional to the
portant to develop features for texture. We present in this number of resolution cells in the image block. It is for this
paper a computationally quick procedure for extracting reason that we call these features quickly computable.
textural features of images and discuss the usefulness of We also investigate the usefulness of textural features for
these features for discriminating between different kinds of categorizing or classifying image blocks from three different
image data. data sets ranging from high-resolution photomicrographs
Early image texture studies have employed autocorrela- to low-resolution satellite imagery. The accuracy of clas-
tion functions [3], power spectra [ 4], restricted first- and sification on multiclass categorization using textural features
second-order Markov meshes [5], and relative frequencies of these data sets was in the range of 80-90 percent.
of various gray levels on the unnormalized image [ 6].
These had some degree of success, but we know little more II, TEXTURAL FEATURES
about texture after finding the results of these experiments Our initial perspective of texture and tone is based on
than before because they did not try to specifically define, the concept that texture and tone bear an inextricable
characterize, or model texture. They only used some general relationship to one another. Tone and texture are always
mathematical transformation which assigns numbers to the present in an image, although one property can dominate
transformed image in a nonspecific way. the other at times. The basic intuitively perceived relation-
Recent attempts to extract textural features have been ships between tone and texture are the following. When a
limited to developing algorithms for extracting specific small-area patch of an image has little variation-i.e., little
image properties such as coarseness and presence of edges. variation of features of discrete gray tone-the dominant
Many such algorithms have been developed and tried on property of that area is tone. When a small-area patch has
special imagery. The subjective parameters (such as the a wide variation of features of discrete gray tone, the
selection of thresholds) associated with the techniques do dominant property of that area is texture. Crucial to this
not enable them to be generalized to imagery other than distinction are the size of the small-area patch, the relative
that processed by the authors. Recently, Rosenfeld and his sizes of the discrete features, and the number of distin-
coinvestigators presented a set of procedures for extracting guishable discrete features. As the number of distinguishable
some textural properties of pictorial data [7]-[9]. In [7] tonal discrete features decreases, the tonal properties will
 ,
612 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, NOVEMBER 1973

predominate. In fact, when the small-area patch is only the 90 degrees

'5d<,,reo,
size of one resolution cell, so that there is only one discrete 135 degrees

m
8
feature, the only property present is tone. As the number of
0 degrees
distinguishable features of discrete gray tone increases
within the small-area patch, the texture property will 2

dominate.
One important property of tone-texture is the spatial Fig. l. Resolution cells 1 and 5 are O" (horizontal) nearest neighbors
to resolution cell•; resolution cells 2 and 6 are 135" nearest neigh-
pattern of the resolution cells composing each discrete bors; resolution cells 3 and 7 are 90° nearest neighbors; and resolu-
tonal feature. When there is no spatial pattern and the tion cells 4 and 8 are 45" nearest neighbors to •· (Note this informa-
tion is purely spatial, and has nothing to do with gray-tone values.)
gray-tone variation between features is wide, a fine texture
results. As the spatial pattern becomes more definite and
involves more and more resolution cells, a coarser texture nearest-neighbor resolution cells themselves. We consider a
results. An excellent set of photographs of different types resolution cell-excluding those on the periphery of an
of texture may be found in [ I3]. image, etc.-to have eight nearest-neighbor resolution cells
The preceding description of texture is, of course, a gross as in Fig. I.
simplification and idealization of what actually occurs. We assume that the texture-context information in an
Discrete tonal features are really quite fuzzy in that they do image I is contained in the overall or "average" spatial
not necessarily stand out as entities by themselves. There- relationship which the gray tones in image I have to one
fore the texture analysis we suggest is concerned with more another. More specifically, we shall assume that this tex-
general or macroscopic concepts than discrete tonal features. ture-context information is adequately specified by the
The procedure we suggest for obtaining the textural matrix of relative frequencies P,; with which two neighbor-
features of an image is based on the assumption that the ing resolution cells separated by distance d occur on the
texture information on an image I is contained in the overall image, one with gray tone i and the other with gray tone j.
or "average" spatial relationship which the gray tones in Such matrices of gray-tone spatial-dependence frequencies
the image I have to one another. More specifically, we shall are a function of the angular relationship between the
assume that this texture information is adequately specified neighboring resolution cells as well as a function of the
by a set of gray-tone spatial-dependence matrices which are distance between them. Fig. 2 illustrates the set of all
computed for various angular relationships and distances horizontal neighboring resolution cells separated by distance
between neighboring resolution cell pairs on the image. All I. This set, along with the image gray tones, would be used
of our textural features are derived from these angular to calculate a distance I horizontal gray-tone spatial-de-
nearest-neighbor gray-tone spatial-dependence matrices. pendence matrix. Formally, for angles quantized to 45°
intervals the unnormalized frequencies are defined by
Gray-Tone Spatial-Dependence Matrices
P(i,j,d,0°) = #{((k,/),(m,n)) E (L, x L,)
Suppose an image to be analyzed is rectangular and bas
N, resolution cells in the horizontal direction and N, x (L, x L,) I k - m = 0, II - nl = d,
resolution cells in the vertical direction. Suppose that the
gray tone appearing in each resolution cell is quantized 1 to l(k,I) = i, l(m,n) = j}
N, levels. Let L, = {1,2,- · · ,N,} be the horizontal spatial
P(i,j,d,45°) = #{((k,/),(m,n)) E (L, x L.)
domain, L, = {l,2,- · · ,N,} be the vertical spatial domain,
and G = {1,2, · · · ,N,} be the set of N, quantized gray x (L, x L,) I (k - m = d, I - n =- d)
tones. The set L, x L, is the set of resolution cells of the
image ordered by their row-column designations. The image or (k - m = -d, I - n = d),
I can be represented as a function which assigns some gray
l(k,I) = i, l(m,n) = j}
tone in G to each resolution cell or pair of coordinates in
L1 X Lx; /: L1 X Lx --+ G. P(i,j,d,90°) = #{((k,l),(m,n)) E (L, x L,)
An essential component of our conceptual framework of
texture is a measure, or more precisely, four closely related x (L, x L,) 11 k - ml = d,
measures from which all of our texture features are derived. I - n = 0,l(k,I) = i,l(m,n) = j}
These measures are arrays termed angular nearest-neighbor P(i,j,d,135°) = #(((k,l),(m,n)) E (L, x L,)
gray-tone spatial-dependence matrices, and to describe
these arrays we must emphasize our notion of adjacent or x (L, x L,) I(k - m = d, I - n = d)
or (k - m = -d, I - n = -d),
1
Variations in lighting, lens, film, developer, and digitizer usually
introduce monotonic transformations of the "true" image gray-tone l(k,I) = i, l(m,n) = j} (I)
values. Under these conditions, we would want two images of the same
scene, one being a gray-tone monotonic transformation of the other
to produce the same features. Image nonnalization by equal-prob-
where # denotes the number of elements in the set.
ability quantizing guarantees that images which are monotonic trans-
formations of one another produce the same results [14], [15]. An Note that these matrices are symmetric; P(i,j; d, a) =
equal-probability quantizing algorithm is given in Appendix III. PU, i; d, a). The distance metric p implicit in the preceding
 HARALICK et al.: TEXTURAL FEATURES FOR IMAGE CLASSIFICATION 613

Grey Tooe
(1,1) (1,2) (J,3) (1,4) 0 1 2 3
0 0 I I o l(O,O) 1 (0, I) l(o,2) l(0,3)
(2, 1) (2,2) {2, 3) (2,.0
Ly= ll,2,3,4! I I #(I ,OJ 1(1, I) #(1 ,2) l(J ,3)
0 0
(3, 1) (3,2) {3,3) (3,4)
LX = !1,2,3,4\ 0 2 2 2 Grey 1(2,0) 1(2, I) 1 (2,2) #(2,3)
(4, l) (4,2) (4,3) {4,4) 2 2 3 3 Tooe 3 #(3,0) #(3, J) 1(3,2) #(3,3)'

(a) (b)
!ik,1),(m,n~ ~ jLyxlxJ xjl xl"J I k-m>= O, ~-nl = 1)
MO,l),{1,2» , ((1,2),(1, 1{, Jl,2),(1,3)) , kt,3), (1,2)), rf' ,rf'
~1,3),(1,4)), tI,4),(1,3)), ~2, l),i2,2)], ~2,2),(2, 1)),
t2,21,(2,3J), k2,3J,t2,2i). [<2,3),t2,4)), ((2,4J,(2,3JJ,

((3, n, (3,2J) • 2J,'(3, 1 ♦ , ~3,2>:t3,3~ , (C3,3J, p,2>J ,

kJ, (c) (d)
f3,3),(3,4~, p,4),{3,3♦, j(4, 1),(4,2~, k4,2),(4,I ♦,

i; !)
t4,2),(4,3j, t4,3),(4,2~, ~4,3),(4,4~, [(4,4),(4,3))]

Fig. 2. Set of all distance 1 horizontal neighboring resolution cells

on 4 x 4 image. J>~plDc (:

equations can be explicitly defined by (e) (f)

Fig. 3. (a) 4 x 4 image with four gray-tone values 0-3. (b) General
p((k,l),(m,n)) = max {lk - ml, 11 - nl}. form of any gray-tone spatial-dependence matrix for image with
gray-tone values 0-3. :ft: (i,j,) stands for number of times gray tones
Consider Fig. 3(a), which represents a 4 x 4 image with i and j have been neighbors. (cHf) Calculation of all four distance
four gray tones, ranging from O to 3. Fig. 3(b) shows the 1 gray-tone spatial-dependence matrices.
general form of any gray-tone spatial-dependence matrix.
For example, the element in the (2,1) position of the most of the images we have processed, the computations
distance I horizontal P 8 matrix is the total number of times have been carried out on a small digital computer (PDP
two gray tones of value 2 and I occurred horizontally 15/20 with 12 kwords of core and two DEC tape drives).
adjacent to each other. To determine this number, we count
the number of pairs of resolution cells in R8 such that the Textural Features Extracted from Gray-Tone Spatial-
first resolution cell of the pair has gray tone 2 and the Dependence Matrices
second resolution cell of the pair has gray tone I. In Fig. Our initial assumption in characterizing image texture is
3(c)-3(f) we calculate all four distance I gray-tone spatial- that all the texture information is contained in the gray-tone
dependence matrices. spatial-dependence matrices. Hence all the textural features
If needed, the appropriate frequency normalizations for we suggest are extracted from these gray-tone spatial-de-
the matrices are easily computed. When the relationship is pendence matrices. The equations which define a set of 14
nearest horizontal neighbor (d = I, a = 0°), there will be measures of textural features are given in Appendix I.
2(N, - I) neighboring resolution cell pairs on each row Some of these measures relate to specific textural charac-
and there are N, rows, providing a total of 2N,(N, - I) teristics of the image such as homogeneity, contrast, and
nearest horizontal neighbor pairs (see Fig. 3). When the the presence of organized structure within the image. Other
relationship is nearest right-diagonal neighbor (d = I, measures characterize the complexity and nature of gray-
. a = 45°) there wiH be 2(N, .,. I) 45° neighboring resolution tone transitions which occur in the image. Even though
cell pairs for each row exoept-the first, for which there are these features contain· information about the textural
none, and there are iv, !Ows, This provides a· total of characteristics of the image, it is hard to identify which
U,N, - l)(N, - I) nearest ·nght-diagonal neighbor pairs. · specific textural characteristic is represented by each of
By symmetry there will be 'l:N,(N, - I) nearesC'Vertical these features.
neighbor pairs and 2(N, - O(N, - I) nearest left-diagonal For illustrative purposes, we will define 3 of the 14
neighbor pairs. After the number of neighboring resolution textural features in this section and explain the significance
cell pairs R used in comp.uting a particular gray-tone of these features in terms of the kind of values they take on
spatial-dependence matrix is ,obtained, the matrfx is for two images of disttnctly different textural characteristics
·normalized by dividing each entry in .the matrix by R. · The features we consider are
It is appropriate at this point to comment on the com-

l1=L L (P(i,j))'
putational aspects of obtaining the gray-tone spatial-de- N, N,
pendence matrices. The number ef operations required to
i=I j=l R
- (2)
process an image using our procedure is directly propor-
tional. tQ the number <>f resolution cells n presen,t in the
· image. In comparison, the number of operations .att of the (3)
order of n log n if one wishe:i to use Fourier or Hadamard
transform to extract texture information. Also, to compute
N,
the entries in the gray-tone spatial-dependence matrices,
one needs to keep only two lines of image data in core al a
/3
L
i I ·- I
[ijP(i,j)/R] - µ,µ, E (4)
time. Thus no severe storage constraints are imposed. For
614 IEEE lltANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS. NOVEMBER J973

IJ-_-• '.'-:--.1~..-.~--:--•l-••-..;~1-~'.'"•+'.'-,::.'.4'"'-::;.,,-,..,..::;H"C.,~'.'--
flffl♦ fH • 0"1111 ff n•Ht
.....
. . . ..
.,.,i,;xn ■
.,
.......................!
~•1n•1••+t.,+ ♦'♦ "'lln....... ttW . . H++.afl111ftltefl..... +
..
I. . . u,■ ••···I ♦ ··~····•-H Hll=
Jb_.. hHNIUhH••aw11111an11 ..........ft1:&111:uwww11
l[JIJ(_j[ + ♦ ,.,,, . . . . . . ¥111111:u·
J ♦i ♦♦ il ♦ iiAl ♦ +++h.iiiiiiiii ♦fiiilf=:+UiiiHW.Wiffiit.U
I ♦I ..... u ..................... Hl¥z t l+ ♦ +=·• ■lll=="'il'INll ■lllllltl
H
iffHihHUh+iihilli=,.liliiii.u+I:ilHIWh=sMiiiili-u-1{
! . . . .h:1--Ulll!H f=<HflillllHIIUl-1♦♦'.0:Ml[ ♦ :iJ:t==ihttlNIIII[ • 1 I!
UAAi ♦HIWiiWTiiih:WH_iirii■ dUhtli=HiiifiWiii •• I
Jtt+ ♦ .. U,. '.ll ♦ + ♦ +H11•=1
•' .. _••.• •II ♦ :l!Jt ♦,.,.=+u,ut11~+11 ·' .I
r ••• • "'··-··-···· - , ...... ~ ........ - ,...,,,... , .......,, ,- •• 1
'"••••'••••••••,, ,.• •••, ,• ••• •• ••••,,•"\II l=ttfHl ■ =t 111 •• .•
l ., • •• • • . . . . . •)i+f ... , •••• t
I , • • • ••• • . • •.•, •••••• - •••• •• ,· •• •• ·•• ••• •• •. •••. I
i _ • , , .. ,. .. ,,, .. ,... • .. , .. ,, .. •• I •• I
t·, -·· •••• -. ·....... ••••••••·•••• . . . . . . . . •,·~·· •1•1•., ... 1
z. .......... -· .............. ••.-~•- ... ,,·.,;, 2
. ..... ...... . ..... ..
T,

1.-..-; .......... ~.••.~

.
• ., .,, ._ ,
. ......... ...•·•ll ••h ..

•.·,~--. ...
1................... - .,,., ...... , ••••- ••• I
. ............... .

Grassland Water Body

Angle ASM Contrast Correlation ASM Contrast Corre lat ion
ff' .0128 3.048 .8075 . 1016 2 .153 .7254
45° .0080 4.011 .6366 .0771 3.057 .4768
90° .0077 4.014 .5987 .0762 3.113 .4646
135° .0064 4.709 .4610 .0741 3.129 .4650
Avg. .0087 3.945 .6259 .0822 2.863 .5327
(a) (b)
Fig. 4. Textural features for two different land-use category images.

whereµ:," µ 1 , ux, and a1 are the means and standard devi- below the images in Fig. 4 shows the usefulness of the ASM
ations of the marginal distributions associated with P(i,j)/R, feature as a measure of the homogeneity of the image.
and R is a norma1izing constant. The contrast feature / 2 is a difference moment of the P
Fig. 4 shows the digital printout of two 64 x 64 image matrix and is a measure of the contrast or the amount of
blocks taken from a satellite picture over the California local variations present in an image. Since there is a large
coastline (NASA ERTS Image no. 1002-18134). The image amount of local variation present in the image of Fig. 4(a)
shown in Fig. 4(a) is a representative sample of grasslands compared to the image shown in Fig. 4(b), the contrast
and Fig. 4(b) is a representative sample of water bodies in feature for the grassland image has consistently higher
the area. The values of the features / 1 , / 2 , and / 3 obtained values compared to the water-body image.
from gray-tone spatial-dependence matrices fo'r distance The correlation feature / 3 is a measure of gray-tone
d = I, are shown below the images in Fig. 4. linear-dependencies in the image. For both the images
The angular second-moment feature (ASM) / 1 is a shown in Fig. 4, the correlation feature is somewhat higher
measure of homogeneity of the image. In a homogeneous in the horizontal (0°) direction, along the line of scan. The
image, such as shown in Fig. 4(b), there are very few dom- water-body image consists mostly of a constant gray-tone
inant gray-tone transitions. Hence the P matrix for this value for the water plus some additive noise. Since the noise
image will have fewer entries of large magnitude. For an samples are mostly uncorrelated, the correlation features
image like the one shown in Fig. 2(a), the P matrix will have for the water-body image have lower values compared to
a large number of small entries, and hence the ASM feature the grassland image. Also, the grassland image has a con-
(which is the sum of squares of the entries) in the P matrix siderable amount of linear structure along 45° lines across
will be smaller. A comparison of the ASM values given the image, and hence the value of the correlation feature is
HARALICK et al.: TEXTURAL FEATURES FOR IMAGE CLASSIFICATION 615

higher along this direction compared to the values for 90°

and 135° directions. More examples of the significance of
some of the textural features are presented in Appendix IL
The various features which we suggest are all functions
of distance and angle. The angular dependencies present a
special problem. Suppose image A has features a, b, c, and
d for angles 00, 45°, 90°, and 135°, respectively, and image
B is identical to A except that B is rotated 90° with respect (a) (b)
to A. Then B will have features c, d, a, and b for angles 0°,
45°, 90°, and 135°, respectively. Since the texture context of
A is the same as the texture context of B, any decision rule
using the angular features a,b,c,d must produce the same
results for c,d,a,b. To guarantee this, we suggest that the
angnlarly dependent features not be used directly. Instead
we suggest that two functions of a, b, c, and d, their average
and range (which are invariant under rotation), be used as (c)
inputs to the classifier.

III. APPLICATIONS OF 'fExTURAL FEATURES FOR IMAGE

CLASSIFICATION
In this section we present the results of our studies on
the usefulness of the textural features for categorizing
images. Three data sets were used in our study. These data
sets were extracted from photomicrographs of different (d) (e)
rocks, from aerial photographs of man-made and natural Fig. 5. Samples of photomicrographs of sandstones. (a) Dexter-L.
(b) Dexter-H. (c) St. Peter. (d) Upper Muddy. (e) Gaskel.
scenes, and from high-altitude satellite pictures of the earth.
A brief description of the data sets and classification al-
gorithms used and the results of classification experiments
15/20 computer is interfaced to the IDECS and the system
will be presented. For further details the interested reader
is capable of performing various functions of image pro-
is referred to [16]-[18].
cessing [25]. The IDECS is capable of processing up to
four image inputs. Three flying spot scanners can input
Data Set Description
transparencies up to a maximum size of 4 x 5 in. A vidicon
1) Photomicrographs of Sandstones: Identification of the scanner accepts larger transparencies as well as printed
types of rocks (sandstones) present in an underground images. Multi-image inputs are electronically registered and
reservoir of crude oil is important in petroleum production combined into a format more suitable for interpretation.
studies; hence we undertook a study to identify different Processed images are displayed on a color display unit
types of rocks from the textural characteristics of their and/or a black and white monitor. A 24-channel disk is
photomicrographs. The data set consisted of 243 image synchronized with the system and used for storage of
blocks of size 64 x 64 and the gray tones of the images were processed digital images. Either manually or under com-
equal-probability quantized into 16 levels. Details of the puter control, the IDECS/PDP is able to perform image
algorithm used to perform this quantization are given in enhancement, video level slicing, area integration, category
Appendix III. There were five sandstone categories in the discrimination, display of histograms and scatograms,
data set, and samples of the photomicrographs are shown in quantizing, and similar image-processing operations. The
Fig. 5. The textural features for the 243 samples were algorithm used for quantizing the aerial photographic data
calculated from distance 1 gray-tone spatial-dependence set digitized the images into eight levels in such a way that
matrices. the areas of the different levels (in each image) were equal.
2) Aerial Photographic Data Set: This data set consisted Each image was digitized into a 20 x 50 array. The eight
of 170 images belonging to eight categories. These 170 categories were residential old (RSOLD), residential new
images were manually sampled from 9 x 9-in 1 : 20 000 (RESNU), LAKE, SWAMP, MARSH, URBAN, RAIL,
aerial photography negatives provided by the Army En- and SCRUB or WOOD (SCROD). The first six categories
vironmental Topographic Laboratories, Fort Belvoir, Va. contained 20 samples each, the RAIL and SCROD cate-
[17]. The digitization of the sampled images was done on gories contained 10 and 40 samples, respectively. Samples
the Image Discrimination, Enhancement, and Combination of the images are shown in Fig. 6.
System (IDECS)/PDP. The IDECS is an analog-digital near 3) Data Set Derived from Satellite Imagery: A variety
real-time processing system that has been developed at the of satellite imagery is currently used for the remote sensing
University of Kansas Center for Research, Inc. A PDP of Earth resources. The most commonly used satellite
616 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNEI'ICS, NOVEMBER 1973

(a) (bl (c)

(d) (e) (f)

(g) (h) (i)

Fig. 6. Samples of aerial photographic data set. (a) RSOLD. (b)
RESNU. (c) LAKE. (d) SWAMP. (e) MARSH. (f) URBAN.
(g) RAIL (h) SCRUB. (i) WOOD. (SCRUB and WOOD are
classified together as SCROD.) Fig. 7. Satellite picture of Monterey Bay, Calif.

imagery is the multispectral scanner (MSS) imagery, which a reasonably large number of samples, we used a piecewise
consists of a set of images of a scene, where each image in linear-discriminant function method. The data sets were
the set is created by scanning the radiance of the scene partitioned arbitrarily into training and test sets. The
from the satellite in a narrow band of wavelengths. The classification algorithm was developed using the samples in
MSS operates in wavelengths ranging from 0.3 to I 6 µm, the training set and it was tested on the samples in the test
and up to 24 scanners have been used in the past. set.
The data set used in our study was derived from a .high- A variety of algorithms have been discussed in the pattern-
altitude four-channel multispectral satellite image taken recognition literature for obtaining piecewise linear-dis-
over the California coastline. The image in one of the four criminant functions for pattern classification [19]-[22]. In
MSS bands is shown in Fig. 7. The ground area covered by a widely used algorithm, the pattern space is partitioned
the image is about 14 400 mi 2 and the size of the digital into a number of regions using a set of hyperplanes (decision
image is 2340 x 3200. Out of this large image, a subset of boundaries) whose locations are determined by the sample
624 contiguous image blocks of size 64 x 64 was taken patterns. Each region is dominated by sample patterns of a
and used in our study. There were seven land-use categories particular category. When a new pattern is presented for
in the image; these are coastal forest, woodlands, annual identification, it is assigned a category depending on the
grasslands, urban areas, small and large irrigated fields, and region to which it belongs. If the new pattern Xis located in
water bodies. The land-use category for each 64 x 64 a region dominated by sample patterns of category ci, then
image block was obtained by interpreting the color com- X is classified as coming from category ci.
posite of the multispectral image set. The textural features For the multicategory problem involving NR categories,
were computed from the distance I gray-tone spatial- a total of NR(NR - 1)/2 hyperplanes are used to partition
dependence matrix of the second MSS band image in the the pattern space. These hyperplanes are defined by a set
multispectral image set. of weight vectors Wii• i = 1,2,· · ·,NR, j = 1,2,· · ·,NR,
Since the ground truth category information was derived j > i, which separates the sample patterns belonging to the
from the color composite image, we used a set of spectral ith and jth categories. A regression type algorithm given in
(color) features in addition to the textural features for [19, ch. 4] was used to obtain the weight vectors. After the
categorizing these images. The spectral features consisted location of the hyperplanes are determined, the classifica-
of the mean and standard deviation of the gray-tone values tion of new patterns is done as follows. For each category
of the 64 x 64 image blocks in each of the four spectral c,, the number of hyperplanes V, which give a positive
bands. response when the new pattern X is presented are deter-
mined using
Classification Algorithms
I) Piecewise Linear Discriminant Function Method: For
v = 2' IW.;TZI + w,/z 1,2,- · · ,N • (5)
categorizing image blocks in data sets I and 3, which have ' i.:!
J .O
2JW,/ZI '
HARALICK et al.: TEXTURAL FEATURES FOR IMAGE CLASSIACATION 617

where Z is the augmented pattern vector obtained by adding TABLE I

C.0NTINGENCY TABLE FOR THE CLASSIFICATION Of PHOTOMICROGRAPHS
a component of value l to X, i.e., OF SANDSTONES

ASSIGNED CATEGOIY Tot<il

(6) Oe><ter-L St. Peter Go>k,I
Dexier-H
"-
M"""y

X is assigned to category c; if V; = max { V,}. If there is a > Dexter-L 29 0 I 0 0 30

0

-
Dexter-H 0 15 0 0 0
tie between categories cm and en, then X is assigned to cm if
"3 St. Peter 2 0 22
• 0 ,.15
Wmnrz ~ 0, or to en if Wm/Z < 0. Several modifications Upper
0 0 4 II 0 15
of the linear-discriminant function method and a multitude ~ Goskel 0 0 0 0 12 12
of other classification procedures may be found in the TOTAL ,, 15 27 15 12 100
references cited.
2) Min-Max Decision Rule: For the aerial photograph Number of samples in test set = 100; number of samples in training
data set which had a small number of samples per category set = 143; overall accuracy of classification of test set = 89 percent.
we used a min-max decision rule for categorizing the images
based on their textural features. The decision boundaries TABLE II
C.0NTINGENCY TABLE FOR THE CLASSIFICATION OF THE AERIAL
were obtained by using all but one of the samples in the PHOTOGRAPHIC DATA SET
data set, and the sample which was left out was then clas-
ASSIGNED CATEGORY
sified. The procedure was repeated for all the samples in TOTAL
kSOLD RESNU IAKE SWAMf MARS!' URBAN RAIL SCROD
the data set to obtain the overall accuracy of the classifica-
RSOLD 17 0 0 0 I I 0 I 20
tion scheme. This method of "leaving one sample out" is RESNU I 15 0 0 I 2 I 0 20
~
widely used by investigators in the pattern-recognition field 0 IAKE 0 0 J9 0 0 0 0 I 20
when the data set has a small number of samples. The 8 SWAMP 0 0 0 19 I 0 0 0 20
< 0 0 0 0 0 0
decision rule is described as follows. u MARSH 12
' 20

~ RAIL
URBAN 15 0 0
Decision rule: If bnk and an,. define the minimum and
2
'I
0 0 0
I
20
0 0 0 0 5
' JO
maximum values of the uniform distribution, (x 1 ,x2 , • • • ,xN)
is assigned to category k if and only if
SCROD
TOTAL 20
0 0
J9
0
J9
0
19
2

17
0
19
0
6
"
51
40

170

b., :,; Xn S an1<• n 1,2,···,N 140 out of 170 images or 82.3 percent of the images were correctly
classified.
N N
I
n (an,.
n=l bn,.)
:?:TT
n"" 1 (anj b,,)
(7)
2) Aerial Photographic Data Set: A min-max decision
for all J such that bn1 S Xn S anj• n = 1,2, · · · ,N. If there rule was used for the classification of 170 images into eight
exists no k such that b,,k S xn S a,,", n = 1,2, · · · ,N, then categories. The processing was done as follows. Four gray-
(x 1 ,x 2 , • · • ,x,) is assigned to category k if and only if tone spatial-dependencies (for four directions) were com-
f min {Ix, - a.,l,lx, - b,.I)
puted for each image. Eleven textural features (/1- /11 as
defined in Appendix I) were computed, yielding a measure-
n"' 1 (a,,k - b,,,.) ment vector of 4 x 11 = 44 components for each image.
:?: f min {Ix, - a,,l,lx, - b,,I) , j = 1,2, ... ,K (8) Computation of the mean, range, and deviation of each
feature over the four directions reduced the dimensionality
n= 1 (an 1 - bn)
to 33. The minimum and maximum statistics were computed
where K is the number of categories. for 169 images and the min-max decision rule was used to
classify the image which was left out. This procedure was
Results of Image Classification Studies repeated 170 times to classify all the images. Out of these
1) Photomicrograph Data Set: A set of 8 variables, com- 170 images, 140, or 82.3 percent of the images, were correctly
prising the mean and variance of the textural features/1 , /2 , classified. The contingency table is given in Table II.
/ 3 , and/9 (for definitions of these features, see Appendix I), 3) Satellite Imagery Data Set: The piecewise linear-
computed from distance I gray-tone spatial-dependence discriminant function method was used for identifying the
matrices, was used as an input to the classifier. The data land-use category of each image block in this data set. The
set was arbitrarily divided into a training set of I 43 samples input variables to the classifier consisted of the mean vari-
and a test set of 100 samples. Piecewise linear-discriminant ance of the four textural features (/1 , /2 , /3 , and/9 obtained
functions for separating the patterns belonging to the five from the distance I gray-tone spatial-dependence matrices)
sandstone categories were derived using the samples in the and eight spectral features (comprised of the mean variance
training set. The contingency table for classifying the of the image gray-tone values) in each of the four spectral
samples in the test set is shown in Table I. The overall bands.· The 624 samples in the data set were divided arbi-
accuracy of the classification scheme on the test set (ob- trarily into a training set of 314 samples and a test set of
tained by summing the diagonal elements in the con- 310 samples. The classifier was trained on the 314 samples
tingency table and dividing by the total number of samples in the training set, and each sample in the test set was
in the test set) was 89 percent. assigned to one of seven possible land-use categories. The
 618 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, NOVEMBER. 1973

TABLE III
CoNTINGENCY TABLE FOR LAND-USE CLASSIFICATION OF SATELLITE
its tones are dark and heavy. Most people could easily
IMAGERY make the observation that the texture on the two images is
ASSIGNED CATEGORY the same. For a machine to find that the textures are the
IC005tal Woodlands ,e.nnuol same, either the images must be probability quantized and
forest G<=- ""=
Area
Lma<
lrrigotPd
Small
Irrigated
Water otol
the features computed from the probability quantized
'""" Fields Fielck
images (which are invariant under monotonic gray-tone
Coastal 23 I 2 0 0 0 I 27
>
forest transformations), or the features themselves must be in-
0 17 10 0
00 Woodlands I 0 0 2B
variant under monotonic gray-tone transformations. Of the
Annual I J 109 I I 0 0
5 Graulands
Urbon Area 0 J 10 13 0 0 0
"'26 textural features described in Appendix I, the angular
second-moment, the entropy, the sum entropy, the difference
~ large
I 2 6 0 37 2
Irrigated field,
Smoll 0
0
" entropy, the information measure of correlation, and the
lrrigoted Fields
Water
0
0 0
'
0
0
0
J
0
24
0
0 JI maximal-correlation features have the invariance property.
Total
" 35
We intend to repeat the experiments reported here nsing
" 26
'" " 42 26
.
36 310
these kinds of features. We expect that these features will
provide more generalized results.
Nwnber of training samples = 314; number of test samples = 310; Additional investigations are necessary to determine the
accuracy of classification on training set = 84.0 percent; accuracy of
classification on test set = 83.5 percent. size of the subimage region and the distances which should
be used in computing the gray-tone dependence matrices.
contingency table for the classification of the test samples is Too small a subimage region will not have enough textural
information to separate image categories of interest, while
given in Table III. The overall accuracy of the classifier
on the test samples was found to be 83.5 percent. a large subimage region may have objects belonging to
Classification of multispectral data such as the satellite several different categories. Also, a large subimage region
imagery we have processed is usually carried out using only will increase the storage requirements. The distance which
the spectral features. We attempted a spectral classification must be used in computing the gray-tone spatial-de-
on the satellite imagery data and achieved a classification pendencies may be obtained from the autocorrelation func-
accuracy of 74--77 percent on the test set. This result, com- tion of the image. The distance at which the normalized
pared with the 83.5-percent classification accuracy achieved autocorrelation function of the image becomes too small
can serve as an upper bound on the distance which may be
using a combination of spectral and textural features, shows
that a significant improvement in the classification accuracy used for computing the gray-tone spatial-dependence
matrices.
might result if the textural features are used as additional
inputs to the classifier.
APPENDIX I
TEXTURAL FEATURES
IV. CONCLUDING REMARKS
We suggest a set of 28- textural features which can be
We have described a class of quickly computable textural extracted from each of the gray-tone spatial-dependence
features which seem to have general applicability to many matrices. The following equations define these features.
kinds of image data. The textural features are based on
statistics which summarize the relative frequency distribu- Notation
tion (which describes how often one gray tone will appear p(i,j)(i,j)th entry in a normalized gray-tone spatial-
in a specified spatial relationship to another gray tone on the dependence matrix, = P(i,j)/R.
image). We have used these features in category-identifica-
PxCi) ith entry in the marginal-probability matrix
tion tasks of three different kinds of image data. Identifica-
tion accuracy on independent test sets are 89 percent for obtained by summing the rows of p(i,j),
the photomicrograph image set (five categories of sand- = L7i, P(i,j).
stones), 82 percent for the aerial photographs (eight land- N, Number of distinct gray levels in the quantized
use categories), and 83 percent for the satellite imagery image.
(seven land-use categories). N11 Ng
These initial experimental results are promising. Much I
i= l
and L
j= 1
respectively.
work needs to be done, however, on gray-tone normaliza-
tion on the imagery and the use of features which are N,
invariant under monotonic gray-tone transformations. The
p,(J) = I
i=l
p(i,jJ.
· reason is that i~ one important sense, texture is independent N9 N9
of tone. Two people examining photographs of the same Px+,(k) = L L p(i,j),
i= I j=l
k = 2,3,- · · ,2N,.
texture may actually be seeing two different, though related, i+ j=k
kinds of tones in the texture. One photograph may have Ng N9
been developed such that its tones are light and thin, and Px-y(k) = L L p(i,j), k = 0,1,- · ·, N, - I.
the other photograph may have been developed such that i= I j= I
li-jl=k
HARAUCK et af.: TEXTURAL FEATURES FOR IMAGE CLASSIFICATION 619

Textural Features where HX and HY are entropies of p, and p,, and

I) Angular Second Moment:
HXYI = -r L p(i,j) log {p,(i)p,U)}
i j
f, = LL {p(i,j)} 2
•
i j
HXY2 = - LL p,(i)p,U) log {p,(i)p,U)}.
2) Contrast: i j

f2 = :~0• n
2
{J, it,
jr-Jl=n
p(i,j)} ·
14) Maximal Correlation Coefficient:
f 14 = (Second largest eigenvalue of Q) 112
3) Correlation: where
LL (ij)p(i,j) - µ,µ, . ') _ ~ p(_i,k)p(j,k)
f, ' . Q( l,J - ~
k p,(i)p,(k)
.

where µx, µ1 , ux, and uy are the means and standard devi- These measures of correlation have some desirable proper-
ations of p, and p,. ties which are not brought out in the rectangular correlation
measure f,. For details, see [23], [24].
4) Sum of Squares: Variance Referring to (I) in the text, for a chosen distance d we
have four angular gray-tone spatial-dependency matrices.
f. = LL
i
(i
j
- µJ 2 p(i,j).
Hence we obtain a set of four values for each of the pre-
5) Inverse Difference Moment: ceding 14 measures. The mean and range of each of these
14 measures, averaged over the four values, comprise the
set of 28 features which can be used as inputs to the
f, = 2: 2: I + /
1 J l -
")' p(_i,j).
} classifier. In this set of 28 features some of the features are
strongly correlated with each other. A feature-selection
6) Sum Average:
procedure may be applied to select a subset or linear
2N,
combinations of the 28 features.
f. = L iv,+,Ci).
i=2
APPENDIX II
7) Sum Variance:
ExAMPLFS OF THE SIGNIFICANCE OF SOME TEXTURAL
2N,
f1 = I (i - f,>
2v•• ,<o. FEATURES
i=2 The problem "What do the textural features represent?"
8) Sum Entropy: 2 from a human perception point of view can be a subject for
2N,
a thorough experiment. There are some intuitive expecta-
fa - L Px+,(i) log {P,+,(i)}. tions as to properties represented by some features. For
i=2 example, one might expect the entropy feature to take higher
9) Entropy: values for more complex images. One might expect to be
able to notice some linear dependencies in images with high
f. = - L L p(i,j) log (p(i,j)). values of the correlation feature. As a beginning for an
i j
investigation of this problem we used an aerial photography
JO) Difference Variance:
data set that consisted of 12 categories, 6 samples in
/ 10 = variance of Px-r each [17].
We ordered the categories in the following manner. We
JI) Difference Entropy:
computed the range and the mode of values that each
Nr,-1
feature takes for each category. The categories were then
f,, = - L
i=O
P,-,(i) log {p,_,(i)).
ordered for each feature according to the modes, and the
ranges were plotted. The prospective of this Appendix is to
12), 13) Information Measures of Correlation:
comment on the ordering of categories by some of the
_ HXY - HXYI features. Figs. 8-10 show plots of the ranges of three of the
f 12 -
max {HX,HY) features, each accompanied by three pictures. These pic-
tures are of units that have small, medium, and large values
f 13 = (1 - exp [ -2.0(HXY2 - HXY)]) 1 l2 of the respective feature. The order of categories of the
HXY = - LL p(i,j) log (p(i,j)) average angular moment feature (ANGMOM) in Fig. 8 is
i j almost the reverse of the order of categories of the average
entropy (ENTROP) in Fig. 9. The three representative
2
Since some of the probabilities may be zero, and log (0) is not pictures show that in both figures the units were ordered
defined, it is recommended that the term log (p + e) (e an arbitrarily according to the complexity of the image. Images with more
small positive constant) be used in place of log (p) in entropy com-
putations. gray levels have lower average ANGMOM and higher
 620 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, NOVEMBER. 1973

R"-IL 1-1 SWAMP 1-•---•-------••I

RSOLD 1-1
H WOO 1------------------l
SW~r,IP l••J
ORCHR 1----1
TWA TTl 1----------------------------------------------l
RSM!=:V 1------ l
SCRU8 1----------------1
U1?9Alf l·••••·•·I
l'IARSH 1------• ------------------------------1
DROAO 1--------1
RAIL J-------------------1
Jl'IARS!f 1····•·····---•·•- J
OIICf.lR !--------- I
SCRUB l·•·····-----1
URBAlf 1---------- l
TWA TR I··•----------------------1 RSCLD 1---1
HWOO I-•-------•------------------------------•... -I IISNEW 1---------1
SWA TR 1--------------------- -------- •••••••••••••• -------- - l flROAD 1-------------------1
I l Ill l I Ill I Ill l l l I II II ll l I l l II II l l l l ll II ll ll I l I l ll J 111111 J l ll l l I I Ill 1111 J 111 l JI I l I Ill I I I J l Ill I l J 111111111 J 11111 I l l l l l l I I l I l l I I l l
l l I I I I I i
0.0.ior &.15, 0,21, ~-.Js-t !1,3(5;,S 0,530 0,702 fl.B?5

RAIL 4 SCRUB 1 SWATR 4 SWATR 4 RAIL 6 DROAD 3

0.037 0.102 0.395 0.418 0.602 0.892

Fig. 8. Range plots and representative pictures for average second Fig. 10. Range plots and representative pictures for average correla-
angular moment. tion.

SWA TR I•----•····•-------•-•---••----------•-------- I I,
H WOO l·•-------------------··•···-------- ---------- -1
T'o'ATR 1------------------------------ ----------- I

S~UB 1--------------------1

DIIOAO 1---••--------- ••••••••• I

URBAN l------ --------- I
RSt/EW

f'\AIISH 1----------------------1

ClRCl!R !----------- --- I

RSOl.fl 1----J

Sli'AMP 1-------1

RAIL 1---------1 Fig. 11. Quantizing algorithm. At kth iteration, F(qt - 1) probability
111 l 11 J I I I l I l 11 l 1 l I l l I II l 1111 11 l Ii I I l l l I l Jl II I l II I l l 11 J 1111 I 1l l J has already been allocated to k - 1 levels and 1 - F(q" - 1)
I I l l probability remains to be allocated. If 1 - F(q" - 1) prob-
l,'J!i'5 ?..519 .~.081 3.iSH
ability is split up equally among remaining K - k + 1 quan-
tizing levels to be allocated, each level would get (1 - F(qt- 1 ))/
(K - k + 1). Since Fis a step function, there is no guarantee that
a q, can be found and that F(q,) - F(q,_,) + (I - F(q,_,))/
(K - k + 1). Hence we look for a qt which is closest to satisfying
the equality.

APPENDIX III
SWATR 6 SCRUB 2 EQUAL-PROBABILITY QUANTIZING
RAIL 6
1.900 2.781 3.700 Let X be a random variable with cumulative probability
function F,. Let Q 1 , the K-level equal-probability quantizing
function for X, be defined by
Fig. 9, Range plots and representative pictures for average entropy.
Q 1 (x) = k, if and only if
average ENTROP. The average ratio (RATIO) in Fig. 10
is lower for natural scenes and higher for man-made scenes.
The three representative pictures in Fig. 10 show corre- lub ( w I F,(w) = k ~ 1
) ,;; x < lub ( w I F,(w) = ; )
spondence between the average RATIO values and the
linear dependency of gray levels in the images. These com- where lub denotes least upper bound.
ments reflect our subjective perception. Much more ex- For any strictly monotonic function g, define the random
perimentation and analysis should be done in this subject. variable Y by Y = g(X). Let Q 2 , the K-level equal-prob-
HARALICK et al.: TEXTURAL FEATURES FOR IMAGE CLASSIFICATION 621

ability quantizing function for Y, be defined by let q, be the smallest number such that
Q2 (y) = k, if and only if
I - F.(.q, ,) + F,(q,_,) - F,(q,)I
k - I) ,,; y < lub ( I K-k+I
lub ( w I F,(w) = ~ w I F,(w) = -~:) .

The following lemma states that the equal-probability

, ; 1~-__F:<~-:) + F,(q,_,) - F(q)I, for all real q.

quantization of X produces a random variable which is Fig. I I illustrates the equal-probability quantizing algorithm.
identical to the equal-probability quantization of Y.
Lemma: Q,(X) = Q2 (Y).
REFERENCES
Proof: Q,(X) = k, if and only if
[1] Special Issue on Digital Picture Processing, Proc. IEEE, vol. 60,
pp. 763-922, July 1972.
1
lub(xlF,(x)=k; ) ,,;X<lub(xlF,(x)= 1)- (9) [2] H. C. Andrews, A. G. Tescher, and R. P. Kruger, ..Image pro-
cessing by digital computer," IEEE Spectrum, vol. 9, pp. 20-32,
July 1972.
(3] H. Kaizer, ..A quantification of textures on aerial photographs,"
Since g is strictly monotonic, (9) holds if and only if Boston Univ. Res. Lab., Boston, Mass., Tech. Note 121, 1955,
AD 69484.
g (tub ( x I F,(x) = k; 1
)) ,,; g(X)
[4] R. Chevallier, A. Fontanel, G. Grau, and M. Guy, "Application
du filtrage optique a l'etude des photographies aeriennes,"
presented at the 11th Congr. Int. Photogrammetric, July 1968.
[5] R. Bixby, G. Elerding, V. Fish, J. Hawkins, and R. Loewe,

< g (1ub (xlF,(x) = ~l). ..Natural image computer," Aeronutronic Division, Philco-Ford
Corp., Newport Beach, Calif., Final Tech. Rep., vol. 1, Puhl.
C-4035, May 1967.
[6] E. M. Darling and R. D. Joseph, .. Pattern recognition from
(10) satellite altitudes," IEEE Trans. Syst. Sci. Cybern., vol. SSC-4,
pp. 38-47, Mar. 1968. •
[7] A. Rosenfeld and E. Troy, "Visual texture analysis," Comput.
Since g is strictly monotonic, we may reverse the order of Sci. Cent., Univ. Maryland, College Park, Tech. Rep. 70-116,
g and lub so that (I 0) holds if and only if June 1970.
[8] A. Rosenfeld and M. Thurston, "Visual texture analysis, 2,"

lub (u(x) I F.(_x) = k; 1

) ,,; g(X)
Comput. Sci. Cent., Univ. Maryland, Coliege Park, Tech. Rep.
70-129, Sept. 1970.
[9] E. B. Troy, E. S. Deutsch, and A. Rosenfeld, "Gray-level manip-
ulation experiments for texture analysis," IEEE Trans. Syst.,
Man, Cybern., vol. SMC-3, pp. 91-98, Jan. 1973.
< lub (u(x) I F,(x) = 1) [10] Proc. Symp. Computer Image Processing and Recognition, Univ.
Missouri, Columbia, Aug. 24--26, 1972.
[11] R. M. Pickett, "Visual analysis of texture in the detection and
recognition of objects," in Picture Processing and Psychopictorics,
(11) B. S. Lipkin and A. Rosenfeld, Eds. New York: Academic
However, Press, 1970, pp. 289-308.
[12] A. Rosenfeld and B. S. Lipkin, "Texture synthesis," in Picture
F,(x) P(X ,,; x) = P(g(X) ,,; g(x)) Processing and Psychopictorics, B. S. Lipkin and A. Rosenfeld,
Eds. New York: Academic Press, 1970, pp. 309-345.
= P(Y ,,; g(x)) = F,(g(x)). [13] P. Brodatz, Textures. New York: Dover, 1966.
(14] R. M. Haralick, K. Shanmugam, and I. Dinstein, .. On some
quickly computable features for texture," in Proc. Symp. Com-
Hence, (11) holds if and only if puter Image Processing and Recognition, Univ. Missouri, Colwn-
bia, vol. 2, pp. 12-2-1-12-2-8, Aug. 1972.
lub(g(x)IF,(g(x)) = k; 1
),,; g(X)
(15] E. E. Hall et al., "A survey of preprocessing and feature extraction
techniques for radiographic images," IEEE Trans. Comput., vol.
C-20, pp. 1032-1044, Sept. 1971.
[16] R. M. Haralick and D. Anderson, "Texture-tone study with
applications to digitized imagery," Univ. Kansas Center for
< lub (o(x)IF,(g(x)) = ~) Research, Inc., Lawrence, CRES Tech. Rep. 182-2, Nov. 1971.
[17J R. M. Haralick, I. Dinstein, K. Shanmugam, and D. Goel,
"Texture-tone study with applications to digitized imagery,"
(12) Univ. Kansas Center for Research, Inc., Lawrence, CRES Tech.
Rep. 182-4, Dec. 1972.
Finally, (12) is equivalent to [18] R. M. Haralick and K. Shanmugam, "Computer classification
of reservoir sandstones," IEEE Trans. Geosci. Electron., vol.
k -
lub ( y I F,(y) = ~ I) ,,; Y < lub (y I F,(y) GE-11, pp. 171-177, Oct. 1973.
[19] K. Fukunaga, Introduction to Statistical Pattern Recognition.
New York: Academic Press, 1972.
[20] K. S. Fu and J.M. Mendel, Adaptive Learning and Pattern Recog-
(13) nition Systems. New York: Academic Press, 1972.
[21] W. Miesel, Computer Oriented Approaches to Pattern Recognition.
so that Q 2 (Y) = k. New York: Academic Press, 1972.
Equal Probability Quantizing Algorithm: Let X be a non- [22] N. J. Nilsson, Learning Machines. New York: McGraw-Hill,
1965.
negative random variable with cumulative probability func- [231 E. H. Linfoot, "An jnformation measure of correlation," Inform.
tion F,. Let Q, the K-level equal-probability quantizing Contr., vol. 1, pp. 85-89, 1957.
[241 C. B. Bell, .. Mutual information and maximal correlation mea-
function for X, be defined by Q(x) = k if and only if sures of dependence," Ann. Math. Statist., vol. 43, pp. 587-595,
qk- i ::; x < q,c We define q0,q1,q2, · · · ,qk in an iterative 1962.
[25] P. N. Anderson et al., "An automated image processing system-
manner. Let q0 = 0. Suppose q,_ 1 has been defined. Then IDECS," Ph.D. project, Univ. Kansas, Lawrence, 1972.