IEEE TRANSACTIONS ON PATTERN AlVALYSlS AND MACHINE INTELLIGENCE, VOL. 18, NO.

8, AUGUST 1996

Texture Features for Browsing and

Retrieval of Image Data
B.S. Manjunathi and W.Y. Ma
Abstract-Image content based retrieval is emerging as an important
research area with application to digital libraries and multimedia
databases. The focus of this paper is on the image processing aspects
and in particular using texture information for browsing and retrieval of
large image data. We propose the use of Gabor wavelet features for
texture analysis and provide a comp~rehensiveexperimental evaluation.
Comparisons with other multiresolution texture features using the
Brodatz texture database indicate that the Gabor features provide the
best pattern retrieval accuracy. An application to browsing large air
photos is illustrated.

837

gested to reduce the redundancy in the representation.

2) An adaptive filter selection algorithm is proposed which can
facilitate fast image browsing.
3) A detailed comparison with the performance of three other
multiscale texture features is provided. Our proposed representation compares favorably in terms of feature computations and retrieval accuracy.
4) Finally, an application to browsing large air photos is demonstrated.

2 TEXTUREFEATUREEXTRACTION
2.1 Gabor Functions and Wavelets
A two dimensional Gabor function g(x, y) and its Fourier transform G(u, v) can be written as:

Index Terms-Digital libraries, image database, content-based image

retrieval, texture analysis, Gabor wavelets.

RETRIEVALof image data based on pictorial queries is an interesting and challenging problem. The recent emergence of multimedia
databases and digital libraries m,lkes this problem all the more
important. While manual image annotations can be used to a certain extent to help image search,, the feasibility of such an approach to large databases is a qu'estionable issue. In some cases,
such as face or texture patterns, simple textual descriptions can be
ambiguous and often inadequate for database search.
The objective of this paper is to study the use of texture as an
image feature for pattern retrieval. An image can be considered as
a mosaic of different texture regions, and the image features associated with these regions can be used for search and retrieval. A
typical query could be a region of interest provided by the user,
such as outlining a vegetation patch in a satellite image. The input
information in such cases is an intensity pattern or texture within a
rectangular window. See Fig. 6 for an example of a texture based
browsing application.
and texture analysis algoTexture analysis has a long hi~~tory
rithms range from using random field models to multiresolution
filtering techniques such as the wavelet transform. Several researchers have considered the use of such texture features for pattern retrieval [181, [191. This paper focuses on a multiresolution
representation based on Gabor filters. The use of Gabor filters in
extracting textured image features is motivated by various factors.
The Gabor representation has been shown to be optimal in the
sense of minimizing the joint two-dimensional uncertainty in
space and frequency [4]. These filters can be considered as orientation and scale tunable edge and line (bar) detectors, and the statistics of these microfeatures in a given region are often used to
characterize the underlying texture information. Gabor features
have been used in several image .analysis applications including
texture classification and segmentation [I], [141, image recognition
[51, [81, [131, image registration, and motion tracking [151.
The main contributions of this paper are summarized below:
1) A simple texture feature representation based on Gabor
features is proposed, and a filter design strategy is sug-

The authors are with the Depavtrnent of Electrical and Computer Engineering,
University of California at Santa Barbara. Santa Barbava, C A 93106-9560.
E-mail: rnanj@ece.ucsb.edu,wei@iplab.ecc~.ucsb.edu.
Manuscript received Dec. 36,1994. Recommended for acceptance by R. Picard.
For information on obtaining reprints of this article, please send e-mail to:
transpami@computer.oyg, and reference I E E E C S Log Number P96055.

where o,, = 1/2zo, and o,

1/2zoy. Gabor functions form a

complete but nonorthogonal basis set. Expanding a signal using

this basis provides a localized frequency description. A class of
self-similar functions, referred to as Gabov wavelets in the following
discussion, is now considered. Let g(x, y) be the mother Gabor
wavelet, then this self-similar filter dictionary can be obtained by
appropriate dilations and rotations of g(x, y) through the generating function:
g
g,n,(x,y) = a-"~(x',
y'),

x' = a-"(xcose

a > 1, m, n = integer

sine), and y' = a-m(-xsin~+ y cose), (3)

where 0 = nz/K and K is the total number of orientations. The

scale factor a-" in (3) is meant to ensure that the energy is independent of rn.

2.2 Gabor Filter Dictionary Design

The nonorthogonality of the Gabor wavelets implies that there is
redundant information in the filtered images, and the following
strategy is used to reduce this redundancy. Let U , and U , denote
the lower and upper center frequencies of interest. Let K be the
number of orientations and S be the number of scales in the multiresolution decomposition. Then the design strategy is to ensure
that the half-peak magnitude support of the filter responses in the
frequency spectrum touch each other as shown in Fig. 1. This results in the following formulas for computing the filter parameters
o, and o, (and thus oxand o,).

where W = U , and m = 0, 1, ..., S - 1. In order to eliminate sensitivity of the filter response to absolute intensity values, the real
(even) components of the 2D Gabor filters are biased by adding a
constant to make them zero mean (This can also be done by setting
G(0,O) in (2) to zero).

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 18, NO. 8, AUGUST 1996

838

the top 15 retrievals are from the same large image. The performance is measured in terms of the average retrieval rate which is
defined as the average percentage number of patterns belonging to
the same image as the query pattern in the top 15 matches.
We observe that the use of om, feature in addition to the mean
improves the retrieval performance considerably. This perhaps
explains the low classification rate of the Gabor filters reported in
[3]where only the mean value was used. On the average 74.37% of
the correct patterns are in the top 15 retrieved images. The performance increases to 92% if the top 100 (about 6% of the entire
database) retrievals are considered instead (i.e., more than 13 of
the 15 correct patterns are present). Some retrieval examples are
shown in Fig. 2. A detailed comparison with other texture features
is given in Section 4.
Fig. 1. The contours indicate the half-peak magnitude of the filter responses in the Gabor filter dictionary. The filter parameters used are
U,=0.4, U,=0.05, K = 6 , a n d S = 4 .

2.3 Feature Representation

Given an image I(x, y), its Gabor wavelet transform is then defined
to be

where " indicates the complex conjugate. It is assumed that the

local texture regions are spatially homogeneous, and the mean
pmn and the standard deviation om,of the magnitude of the transform coefficients are used to represent the region for classification
and retrieval purposes:

u...

J j(~.~,(xy)~dx~y.
and om,,= J j J (I W ~~, , ( X.

Y)

- i i n , n ) 2 ~ x(6)
d~

A feature vector is now constructed using p,, and om,as feature components. In the experiments, we use four scales S = 4 and
six orientations K = 6, resulting in a feature vector

2.3.1 Distance Measure:

7"'

Consider two image patterns i and 1, and let

and f"' represent
the corresponding feature vectors. Then the distance between the
two patterns in the feature space is defined to be

where

a@,,) and a(o,,) are the standard deviations of the respective

features over the entire database, and are used to normalize the
individual feature components.

2.4 Retrieval Performance

2.4.1 Texture Database
The texture database used in the experiments consists of 116 different texture classes. Each of the 512 x 512 images is divided into
16 128 x 128 nonoverlapping subimages, thus creating a database
of 1,856 texture images. A query pattern in the following is any
one of the 1,856 patterns in the database. This pattern is then processed to compute the feature vector as in (7). The distance d(i, j),
where i is the query pattern and j is a pattern from the database, is
computed. The distances are then sorted in increasing order and
the closest set of patterns are then retrieved. In the ideal case all

Fig. 2. Texture browsing using Gabor features. Examples shown are

some of the difficult patterns to analyze. The average retrieval rate is
shown in the parentheses: (a) Texture browsing interface with all 116
textures and retrievals for D42, lace (50%),(b) D43, swinging light bulb
(54%),(c) D23, beach pebbles (54%),and (d) D91, clouds (25%).The
first image in each of (b)-(d)represents the query image. Note that some
of the incorrect matches actually look quite similar to the query pattern.

In many cases it is desirable to reduce the image processing time

to the extent possible while not seriously affecting the overall performance. One way to reduce the computations is to select the
Gabor filters in a pattern dependent way.

IEEE TRANSACTIONS ON PATTERN AlVALYSlS AND MACHINE INTELLIGENCE, VOL. 18, NO. 8, AUGUST 1996

An insight into the discrimination quality of individual features

can be obtained by considering the average intraclass to inter-class
distance ratio for each of the components. This ratio is about 0.25
on the average, with higher frequency components providing a
better discrimination [16]. In database search one is often interested in finding out how much of the search space can be eliminated by using a particular feature. Suppose we want to keep all
the 15 correct textures belonging to the same pattern in the set of
retrieved images. The average threshold is about 30%, and as in
the previous case, high frequency components have a better discrimination. More details can be found in [161. These experiments
indicate that each feature component can individually provide
useful discrimination measure which can be further improved by
selectively choosing these filters based on the query pattern.
The selection scheme described here uses the spectral information in conjunction with the average database image properties to
select a subset of filters. The pur-pose is to identify salient query
image properties which best diljtinguish it from the database
items. Fig. 3 shows a schematic diagram of this method. The difference between the spectrum of the input image pattern and the
average spectrum provides information about salient spectral
characteristics of the given image. Let

input
pattern

FFT

839

#4
I

selected filters

Fig. 3. An example of the filter, selection strategy.

where <,,p,,,(n,v) is the Fourier transform of the input image pattern. F,,, (u,v) and F,,(tl, v) are the mean and variance associated
with the distribution of Fourier tre~nsformsof all image patterns in
the database. D(u, v) basically is the energy of the difference normalized by the variance associated with each frequency component (u, v). Each filter is evaluated based on the total difference energy within its spectral coverage:

where G,,(u,v) is the frequency response of the filter g,,(x, y).

The larger the value i,,
is, the better the performance of the filter
will be for pattern retrieval. Thus the filters can be ordered based
on
Fig. 3 shows an example where the input pattern has
strong orientation preference which distinguishes the pattern from
much of the database images. Using only the top four filters, one
can retrieve on the average about 50% of the correct patterns in the
top 15 retrieved patterns 1161. Such a strategy may be reasonable in
many image database applications and amounts to a significant
savings in image processing con~putations.Fig. 4 shows some
examples wherein different number of filters are used.
Computing the entire Gabor fe4aturevector of (7) takes 9.3 seconds of CPU time (in MATLAB on a SUN-Sparc20, see Table 2).
Search and retrieval takes about 1.02 seconds. Using only four
feature components requires 2.3 seconds of feature extiaction time
including the adaptive filter selection (0.7 seconds), and about 0.1
seconds for search and retrieval.

c,,.

A detailed comparison with some of the other recently proposed

multiresolution texture image features are made in this section. For
the Gabor feature case, the 24 x 2 component feature vectors are
used. An application to browsing large air-photos is illustrated.

4.1 Comparison with Other Te:~tureFeatures

The comparisons are made with the conventional pyramidstructured wavelet transform (T'WT) features, tree-structured
wavelet transform (TWT) features, and the multiresolution simultaneous autoregressive model (MR-SAR)features. The filter coefficients used for computing PWT are the 16-tap Daubechies or-

Fig. 4. Examples of retrieval using a different number of filters selected

by the adaptive filter selection strategy. The upper-left image in each
block is the query image (D83: woven matting), and the top 15
matches are shown in row scan order with increasing distance. Top:
Results using one filter; bottom: results using two filters. With four filters, all the correct 15 patterns are retrieved for this case.

thogonal wavelets [61 (same as the ones used in [3] for the TWT).
The 128 x 128 image pattern is decomposed into three levels (4 x 3
= 12 bands) of the wavelet transform. The mean and standard deviation of the energy distribution corresponding to each of the
subbands at each decomposition level are used to construct a (12 x 2)
feature vector.
In [3], decomposition of image subbands at each level is based
on energy considerations and this results in a tree structured decompos$ion where different patterns have different structures. For
pattern retrieval applications, it is convenient to have a fixed
structure. A fixed d&compositiontree can be obtained by sequentially decomposing the LL, LH, and HL subbands. The HH band is
not decomposed as this often does not lead to stable features. A
three level decomposition results in 52 (4(1 + 3 + 9)) subbands. As
in the PWT, the mean and standard deviation in each subband are
used to construct a 52 x 2 component feature vector.
The third set of feature used are the MR-SAR model features
[171. Previous work [91, [201 indicate that the MR-SAR features at
levels 2, 3, and 4, provide the best overall performance. At each
level, five parameters are computed to represent the texture, thus
requiring a total of 15 feature components. The Mahalanobis distance is used to compare the feature vectors.
4.1.1 Summary of Comparisons
Table 1 provides a summary of the experimental results. It shows
the retrieval accuracy of the different texture features for each of
the 116 texture classes in the database. The Gabor features give the
best performance at close to 74% retrieval. This is closely followed
by the MR-SAR features at 73%. The TWT features perform mar-

IEEE TRANSACTIONS ON P A T E R N ANALYSIS AND MACHINE INTELLIGENCE, VOL. 18, NO. 8, AUGUST 1996

TABLE 1

AVERAGERECOGNITION
RATE FOR THE 116 TEXTURE
IMAGES IN THE DATABASE.
FROM THE USC DATABASE.
THE D * LABELS INDICATE TEXTURES FROM THE BORDATZ ALBUM [2] AND O* LABELS INDICATE TEXTURES
Average Retrieval Rate%

Average Retrieval Rate%

Gabor

PWT

TWT

MRSAR

99.17
52.92
9458

97.08
32.50
7542

97.50
36.25
58 33

98.75
67.92
49.17

D42
D(3
W

D4 100.00 90.83 67 92
D5 72.92 52.92 52.08
D6 100.00 100.00 l 0 0 m
D7 35.42 21.25 1958
D8 95 00 79.58 74 58
D9 93.75 84.58 77.50
Dl0 85.83 78.75 68.75
Dl1 100.00 7375 80.00

88 33
6333
100.00

D45

Dl
D2
D3

52.08
95.42
61.58
79.17

D46
D47
D-18
D49
D50
D51

98.33

D52

70.42
30.83

75.00
43 75

Dl4 100.00 1W.00 100.00

Dl5 69.17 79.58 57 92

10000
8000

D53
D54
D55

Dl6 100.00 100.00 100.00

Dl7 100.00 100.00 100.00
Dl8 7917 79.17 92.50

100.00

Dl2
Dl3

86.25
42.92

7958
38.75

100.00
69.58

Dl9 80.42 73.33 62.50

D20 100.00 8750 94.58
D21 100.00 100.00 100.00
D22 75.00 82.50 7083
D23 5375 41.25 41.67
D24 85 83 95.83 82.92
D25 88.75 53.75 48.33

8667
100 00
10000
56.67
39.58
90.03
75.83

D26 100.00

88.75

99.17

100.00

1127

36.67

34.58

34.58

10.42

D28

95.42
72.08

8667
60.00

9750
64.58

88.75
67.50

33.75
77 92

23.75
72.50

33.75
57.50

I134 9917
035 9833
D36 49.17
D37 100.00
J>38 4667
D39 3958
D40 52.08
D41 78 75

92.92
82.92
57.08
78.75
31.67
24.17
56.67
68.33

D29
D30
D73

056
D57
D58
D59
DK)
D61
D62
D63
D64
D65
D66
D67

Gabor
50.00
11.25
1250
14.58
94.17
100.00
1917
100 00
87.92
83.75
72.08
100.00
50.83
lM.00
10300
10000
29.58
20.42
5250
13.75
35.83
3117
9458
100 00
96.67
7000

068 100.CO
D69 4250

PWT
59.17
13 75
13 33
22.m
70.42
lOO.00
7108
100.00
5625
91.25
55.42
100.00
56.67
9708
10000
9117
18.33
1083
3000
-17.92
15.00
2117
9000
100.00
9000
53.75

Average Retrieval Rate%

TWT MRSAR
56.67
38.75
11 25
9.58
1192
15 00
25 83
3 33
80.12
92.50
10000
9750
77.08
8667
100.00 100.00
75.83
87.08
9333
90.83
6125
70.00
100.00 10000
-17.50
57.92
99.17 100.00
100.00 1OO.00
100.00 100.00
25.42
27.92
14.17
2333
3792
5000
41.25
-15 83
50.83
-13.75
27.08
35.00
97.92
79.17
98.75 100.00
95.00
68.33
6375
6208

9958 I0000
39.17 44.17

D81
D82
D83
D84
D85
D86
D87
D88
D89
D90
D91

92.08
30.00

68.33
62.08

D94 100.00

91 67

92.08

90.83

87.50
98.33
37.08

65 CO
7750
29.17

92.50
94.17
39.58

87.50
99.17
37.92

52.50
87.08

52.08
71.67

65.00
53.33

72.92
74.58

DIM
Dl07

58.75
53.33
56.U
54.58
63.33
14.17
52.50

6500
5125
7250
59.17
50.00
55.83
59.58

63.75
51.25
73.75
66.67
44.58
54.58
60.83

62.92
56.25
66.67
48.75
53.33
51.67
5542

Dl08

3750

18.75

29.58

32.92

DIG9
Dl10

78 75

71.75

76.67

76.25

87.92
90.83
61.67

7875
9042
50.42

75.42
62.92
58 33

62.92
91.67
63 33

D95
D96
D97
D98
Dl00
Dl01
Dl02
Dl03
DIM
Dl05

79.17
72.50
79.17

87.50
72.92

TWT MRSAR
95.83
94.17
100 00 10000
99.58
99.17
100.00
99.58
100.00
99.17
71.67
95.00
82.92
92.92
51.67
45 83
2625
10.00
3167
47.08
16.25
27.50

92.50
38 75

D92
D93

100.00
36.25

Gabor PWT
100.00 90.83
100.00 IOO.00
100.00 98.75
100.00 100.00
99.58 9667
91 67 6083
9958 9208
41.67 48.75
21.25 2208
34.58 19.58
25.42 12.92

50.12
92.50

D70
D71
D72

1917
4292
-17.50

15.42
15.83
18.75

57.50
68.75
75 83

98.33

100.00

D73

66.67

51.67

5792

5708

0 1 100.00 100.00 10000

9208

68.33
30.00

84.17
3167

79 58
29 17
27.92
67.50
52.08

99 58
4208
6000
41 67
58.33

D71 78.75 85.00 8500

D75 95.12 86.67 9375
D76 99 17 96.25 95.83
D77 10000 100.00 1000i)
D78 97.50 93.33 85.42
,779 100.00 100.00 10000
1380 100.00 85 83 85 83

94.58
98.75
92.08

0 2 100.00 100.00 '18.75

0 3 98.75 75.42 93.33
0 4 1W.IX) 10000 100.00

100.00
97.50
100.00

100.00
88.33
97.92
'11 67

0 5 86.25 83.33 100.00

0 6 100.00 10200 100.00
0 7 71.Z 68.75 5208

93.33
100.00
57.50

Dill
Dl12

Avg.

74.37

68.70

69.41

73.18

TABLE 2
CPU TIMES(ON A SUN S P A R C 2 0 WITH ONE PROCESSOR) AND FEATURE
VECTOR LENGTHFOR THE VARIOUS TEXTURE
FEATURES.
GABOR FEATURES
ARE COMPUTED
IN M A T L A B AND ALL THE OTHERS ARE WRITTENIN C LANGUAGE.
Gabor Pcatures

pfi";:;

~ u l featuie
l

Peaturc ExWaction
T~mc
Pealure Vector
Lenzlh

Scuihing and
Sortine Time

PWT

TWT

MRSAR

9.3 sec.

2.3 scc

1.3 scc

2.3 sec

34.0 sec

48

8
(4x2)

24
(12x2)

104
(52x2)

15

(2x2)
1.02 sec

0 l sec

ginally better (69.4%) than the PWT features (68.7%).Fig. 5 shows

a graph illustrating this retrieval performance as a function of
number of top matches considered. In summary,
In general, feature components corresponding to higher frequencies have better discriminating performance. However,
decomposing the HH band in the tree-structured representation often leads to a decrease in performance, indicating
that these features are not very robust.
Experiments with different wavelet transforms indicate very
little variation in performance with respect to the choice of
filters [lo].
The marginal improvement of the TWT features comes at
the expense of having a much larger feature vector, which
adds to the overhead associated with indexing and search.

a98 s

1 70 5 1 1

0 70 scc

1
(

It is important to explore different similarity measures for

each of the different sets of features. For example, using the
Mahalanobis distance instead of the Euclidean distance improved the performance from 64% to 73% for the MR-SAR
features. Normalized Euclidean distance worked better for
all the others.
For Brodatz images, best results using the Gabor features were
obtained using four scales and six orientations within each
scale.Note that rotation and scale invariance is not addressed.
Table 2 vrovides the CPU times for feature extraction and sequential search of the database. The computations involving Gabor features are performed in MATLAB where as the other feature
computations are implemented in C language. In terms of feature
comhutation time, the MR-SAR is the most expensive.

IEEE TRANSACTIONS ON PATTERN AIVALYSIS AND MACHINE INTELLIGENCE, VOL. 18, NO. 8, AUGUST 1996

841

Number of top matches considered

Fig. 5. Retrieval performance according to the number of top matches

considered.

4.2 An Image Browsing Example

Query based on texture properties will have many applications in
image and multimedia databases. Here, we describe with an example our current work on incorporating these features for
browsing large satellite images and air photos. This work relates to
the UCSB Alexandria digital library project [21] whose goal is to
create a digital library of spatially indexed data such as maps and
satellite images. Typical images in such a database range from few
megabytes to hundreds of megabytes, posing challenging problems in image analysis and visu;~lizationof data. Content based
retrieval will be very useful in this context in answering queries
such as "Retrieve all LANDSAT images of Santa Barbara which
have less than 20% cloud cover," or "Find a vegetation patch that
looks like this region."
We are currently investigating t.he use of texture primitives to accomplish rapid content based browsing within an image or across
similar images. Fig. 6 shows an example of browsing 5,248 x 5,248
air photos. The original image is analyzed in blocks 128 x 128 pixels and the texture features are computed and stored as image
"meta-data." The user can select any position and use that pattern to
search for similar looking regions. Our current work is on incorporating simple texture based segmentation schemes into this browsing thus allowing arbitrarily shapeoi regions into the analysis.

4.3 Discussions
A Gabor wavelet based texture analysis scheme is proposed and
its application to image databases is demonstrated. A comprehensive performance evaluation of thse method is given using a large
number of textures and a comparison with some of the well
known multiresolution texture classification algorithms is made.
Further, a novel adaptive filter selection strategy is suggested to
reduce the image processing computations while maintaining a
reasonable level of retrieval performance. The experimental results
indicate that these Gabor feature are quite robust. Rotation and
scale invariance is important in nnany applications and our preliminary results on rotation invariant classification 171 using Gabor
features are very encouraging.
Finally, a note on similarity measures. It is widely acknowledged
that this is an important but a difficult problem. Our initial results
using simple hybrid neural network learning algorithms appear
very promising in the context of learning similarity 1111, [12].

This research was partially supported by National Science Foundation grant IRI-9411330 and by NASA under grant number

Fig. 6. An example of browsing a large air photo using Gabor texture

features. Part (a) shows the down-sampled version of the image and
the retrieval results on a vegetation pattern, (b) the query pattern contains a part of the highway and retrievals are all from highway segments, (c) shows the region containing some buildings (center of the
image toward the left), (d) shows one interesting result where the query
contains a number marked on the image (lower left corner), and the top
matches also contain similar patterns.

NAGW-3951. We thank Professor Picard and F. Liu at MIT for

providing the software for the MR-SAR features and C. Fischer at
UCSB for the air photos.

[I] A.C. Bovic, M. Clark, and W.S. Geisler, "Multichannel Texture

Analysis Using Localized Spatial Filters," IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 12, no. 1, pp. 55-73, Jan. 1990.
[21 P. Brodatz, Textures: A Photographic Album for Artists O Designers.
New York: Dover, 1966.
131 T. Chang and C.-C.J. Kuo, "Texture Analysis and Classification
with Tree-Structured Wavelet Transform," lEEE Trans. Image
Processing, vol. 2, no. 4, pp. 429-441, Oct. 1993.
141 J.G. Daugman, "Complete Discrete 2D Gabor Transforms by Neural Networks for Image Analysis and Compression," IEEE Trans.
ASSP, vol. 36, pp. 1,169-1,179,July 1988.
[51 J.G. Daugman, "High Confidence Visual Recognition of Persons
by a Test of Statistical Independence," IEEE Trans. Pattevn Analysis
and Machine Intelligeence, vol. 15, no. 11, pp. 1,148-1,161, Nov. 1993.
161 I. Daubechies, "The Wavelet Transform, Time-Frequency Localization and Signal Analysis," IEEE Trans. Information Theoy, vol. 36,
pp. 961-1,005, Sept 1990.
[71 G.M. Haley and B.S. Manjunath, "Rotation Invariant Texture
Classification Using the Modified Gabor Filters," Proc. IEEE ICIP
'95, vol. I, pp. 262-265, Washington D.C., Oct. 1995.

842

IEEE TRANSACTIONS ON P A T E R N ANALYSIS AND MACHINE INTELLIGENCE, VOL. 18, NO. 8, AUGUST 1996

[8] M. Lades et al., "Distortion Invariant Object Recognition in the

Dynamic Link Architecture," lEEE Trans. Computers, vol. 42, no. 3,
pp. 300-311, Mar. 1993.
191 F. Liu and R.W. Picard, "Periodicity, Directionality, and Randomness: Wold Features for Image Modeling and Retrieval," MIT
Media Lab Technical Report no. 320, Mar. 1995.
[lo] W.Y. Ma and B.S. Manjunath, "A Comparison of Wavelet Features for Texture Annotation," Proc. IEEE Int'l Conf. Image Processing '95, vol. 11, pp. 256-259, Washington D.C., Oct. 1995.
[Ill W.Y. Ma and B.S. Manjunath, "Image Indexing Using a Texture
Dictionary," Proc. SPIE, vol. 2,606, pp. 288-296, Philadelphia, Oct.
1995.
[I21 W.Y. Ma and B.S. Manjunath, "Texture Features and Learning
Similarity," Proc. IEEE Conf. Computer Vision and Pattern Recognition, San Francisco, June 1996.
1131 B.S. Manjunath and R. Chellappa, "A Feature Based Approach to
Face Recognition," Proc. IEEE Conf. C V P R '92, pp.
373-378, Cham*
paign, 11l.,Iune 1992.
[I41 B.S. Manjunath and R. Chellappa, "A Unified Approach to
Boundary Detection," IEEE Trans. Neural Networks, vol. 4, no. 1,
pp. 96-16, Jan. 1993.
[15] B.S. Manjunath, C. Shekhar, and R. Chellappa, "A New Approach
to Image Feature Detection with Applications," Pattern Recognition, Apr. 1996.
1161 B.S. Manjunath and W.Y. Ma, "Texture Features for Browsing and
Retrieval of Image Data," CIPR TR 95-06, July 1995 (available on
www at http://vivaldi.ece.ucsb.edu).
[I71 J. Mao and A. Jain, "Texture Classification and Segmentation
Using Multiresolution Simultaneous Autoregressive Models,"
Pattern Recognition J., vol. 25, no. 2, pp. 173-188,1992.
[18] W. Niblack et al., "The QBIC Project," Proc. SPIE, vol. 1,908, pp.
173-181, Feb. 1993.
[I91 A. Pentland, R.W. Picard, and S. Sclaroff, "Photobook: Tools for
Content Based Manipulation of Image Databases," Proc. ICASSP
'93, vol. V , pp. 161-164,Minneapolis, Apr. 1993.
[20] R.W. Picard, T. Kabir, and F. Liu, "Real-Time Recognition with
the Entire Brodatz Texture Database," Proc. IEEE Conf. C V P R '93,
pp. 638-639, New York, June 1993.
1211 The Alexandria Bgital Library Project, http://alexandna.sdc.ucsb.edu.
A

Retrieving Multispectral Satellite Images

Using Physics-Based Invariant
Representations
Glenn Heaiey and Amit Jain
Abstract-We present a set of algorithms and a search strategy for
the robust content-based retrieval of multispectral satellite images.
Since the property of interest in these images is usually the physical
characteristics of ground cover, we use representations and methods
that are invariant to illumination and atmospheric conditions. The
representations and algorithms are derived for this application from a
physical model for the formation of multispectral satellite images. The
use of several representations and algorithms is necessary to interpret
the diversity of physical and geometric structure in these images.
Algorithms are used that exploit multispectral distributions,
multispectral spatial structure, and labeled classes. The performance
of the system is demonstrated on a large set of multispectral satellite
images taken over different areas of the United States under different
illumination and atmospheric conditions.
Index Terms-Image database,image retrieval, color constancy,
satellite images, color, machine vision, texture, computer vision,
recognition.

MULTISPECTRAL
remote-sensing data are being used for an increasing number of applications in a diverse set of fields including
agriculture, geology, mapping, water resources, and environmental science. The volume of satellite data that is available for
such applications is staggering. A Landsat thematic mapper, for
example, generates seven band images using three visible and four
infrared regions of the spectrum. Even though the Landsat provides coarser spatial resolution than many other remote-sensing
satellites, a single image corresponding to a 170 km by 185 km
region of earth requires over 200 Mbytes of storage and the satellite generates about 5,000 images per week. It has been predicted
that in a few years the amount of data originating from remotesensing satellites will reach a terabyte per day [Z].
~ i v e this
n large volume of data, effective tools for image access
are essential to allow the information in the database to be fully
exploited. Image database systems traditionally access images
using keywords or text associated with the images. Unfortunately,
it is often difficult to assign textual descriptions to images and
consequently
text-based queries often fail. In addition, the task of
.
.
manually annotating the current volume of satellite imagery
would involve a large amount of time and expense.
A recent trend & retrieval from image databases has been to
allow queries based on image content [12], [Ill, [15], [I]. Using this
paradigm, a user can specify a search using image properties such
as shape, color, or texture. In most cases, the user will not specify
numerical values of these properties, but rather will present the
system with example images. The system will compute features
from the example and use these features for database indexing. In
most cases, a detric is defined on these features that is intended to
model perceptual similarity.

The authors are with the Department of Electrical and Computer Engineering, University of California, Irvine, C A 92717.
E-mail: {healey, amit)@ece.uci.edu.
Manuscript received July 5,1995.
Recommended for acceptance by R. W. Picard.
For information on obtaining reprints of this article, please send e-mail to:
transpami@computer.org, and reference IEEECS Log Number P96052.
0162-8828196$05.00 0 1996 IEEE