Vous êtes sur la page 1sur 10

Multiband semifragile watermarking for multi and

hyperspectral images based on iterative tree structured


vector quantization
Jordi Serra, David Megı́as, Jordi Herrera-Joancomartı́, Julià Minguillón
Estudis d’Informàtica, Multimèdia i Telecomunicacions
Universitat Oberta de Catalunya
Rambla del Poblenou 156, 08018 Barcelona, SPAIN

ABSTRACT
In this paper a novel semifragile watermarking scheme for images with multiple bands is presented. We propose to use the
remote sensing image as a whole, using a vector quantization approach, instead of processing each band separately. This
scheme uses the signature of the multispectral or hyperspectral image to embed the mark in it and detects a modification
of the original image, e.g. a replacement of a part of the image into the same image or any other similar manipulation. A
modification of the image means to modify the signature of each point, all the bands simultaneously, because in multispec-
tral images it does not have sense to modify a single band of all those that compose the multispectral image. The original
multispectral or hyperspectral image is segmented in three-dimensional blocks and, for each block, a tree structured vector
quantizer is built, using all bands at the same time. These trees are manipulated using an iterative algorithm until the
resulting image compressed by the manipulated tree satisfies all the imposed conditions by such tree, which represents the
embedded mark. Each tree is partially modified accordingly to a secret key in order to avoid copy-and-replace attacks,
and this key determines the internal structure of the tree and, also, the resulting distortion, in order to make the resulting
image robust against near-lossless compression. The results show that the method works correctly with multispectral and
hyperspectral images and detects copy-and-replace attacks from segments of the same image and basic modifications of
the marked image.
Keywords: Semifragile watermarks, Image authentication, Alteration detection, Multispectral images, Hyperspectral im-
ages, Tree Structured Vector Quantization

1. INTRODUCTION
Multi and hyperspectral imaging has gained an increased attention by the research community in the last few years, since
new uses and applications of this area are often reported. High precision farming, water pollution control or natural
resources management, among others, are well-known uses of remote sensing, see 1 and2 for example. Hyperspectral
images have an inherent high economic value since their acquisition involves expensive mechanisms like plane flights or
satellites. Such an economic value must be preserved when hyperspectral images are used in different applications where
a third party pays for such usage.
Data hiding schemes, such as watermarking, have been proven successful in order to protect images in several ways,
such as resolving authoring disputes, obtaining traceability in illegal image distribution issues or detecting changes in the
images aimed to produce a forged copy, as shown in 3 . Many watermarking algorithms are based on robustly embedding
an imperceptible mark into the image in such a way that the mark can be recovered later in spite of possible manipulations.
Every watermarking scheme can be described in two stages: mark embedding and reconstruction, although the real issue
is where and how place the hidden mark. Intuitively, as much information as possible should be embedded in the original
image to ensure that most of such information will be detected in a manipulated image. Nevertheless, the embedded mark
Further author information: (Send correspondence to J. Serra)
J. Serra: E-mail: jserrai@uoc.edu, D. Megı́as: E-mail: dmegias@uoc.edu,
J. Herrera-Joancomartı́: E-mail: jordiherrera@uoc.edu, J. Minguillón: E-mail: jminguillona@uoc.edu,
Telephone (+34) 93 326 36 00, Fax (+34) 93 356 88 22
should not be perceptible and should not be easily removed by simple manipulations of the digital content, thus a tradeoff
between robustness and perceptual quality must be achieved, see 4 for more details.
In hyperspectral imaging applications, the most useful watermarking schemes aimed to detect changes in the image
are semifragile ones. Semifragile schemes are able to overcome some minor modifications, as those produced by near-
lossless compression, but reveal the existence of traitorous manipulations, such as copy-and-replace attacks, or excessive
information removal by means of cropping or lossy compression.
Most existing semifragile watermarking applications can be applied on monochromatic images, and thus, can be ex-
tended to hyperspectral images by treating several bands independently. However, in this paper, a semifragile watermarking
scheme specially developed for hyperspectral images is presented. This method works with all the bands at the same time
and, therefore, provides uniform protection for hyperspectral imaging applications. The method can be tuned to embed the
mark according to band importance, depending on the content to be protected.
In this paper, we propose to use the multispectral image as a whole, using a vector quantization approach. The original
multispectral (CASI)5 or hyperspectral (AVIRIS) 6 image is segmented in three-dimensional blocks of M ×N ×b, where M
and N determine the spatial resolution of the embedding and detection algorithm, and b is the number of bands employed
for building the tree structured vector quantizers. For each block, a tree is built, and these trees are manipulated using an
iterative algorithm until the resulting image satisfies all the imposed conditions by the built trees. Each tree is partially
modified accordingly to a secret key in order to avoid copy-and-replace attacks, and this key determines the internal
structure of the tree, and and also the resulting distortion, in order to make the resulting image robust against lossy (really
near-lossless) compression.
This paper is organized as follows. In Section 2, the coding of remote sensing images by means of tree structured vector
quantization is reviewed. In Section 3, the watermarking strategy and the mark embedding and mark retrieval processes
are described. Section 4 presents the results obtained with the suggested scheme for the chosen experimental corpus and
analyzes the basic parameters that determine these results. Finally, the most relevant conclusions of this work are drawn in
Section 5.

2. CODING REMOTE SENSING IMAGES


The first problem that one has to face when working with hyperspectral images is their huge size. A typical hyperspectral
image covering a small region of a few kilometers has millions of pixels, and each pixel is represented by more than one
band, depending on sensor type. For example, CASI images use 14 bands for each pixel and AVIRIS images use 224
bands. Moreover, the image size is further increased because high resolution is used for each pixel band, up to 16 bpp
instead of 8 bpp. One of the techniques to reduce the large amount of data inherent to hyperspectral images is to employ a
lossy compression method. Lossy, in opposition to lossless, means that, once the input image has been encoded and later
decoded, the recovered image is not exactly the same as the original input image. A few years ago, lossy compression
seemed unacceptable for some applications, but this is changing as end-users are more willing to accept it 7 .
Lossy methods remove information that is not “important” for image reconstruction. But, what is the meaning of
“important”? This is the key issue in lossy image compression, because the removed information should depend on user
purposes. For instance, it is not the same to compress an image for visualization than for automated classification. Several
criteria for image quality can be defined 8 depending on the desired goal. Preliminary experiments 9, 10 show that it is
possible to achieve high compression ratios without removing critical information for classification purposes, for example,
when only limited classification accuracy is achievable.

I P T Q BPE EC I*

Figure 1. Encoder block diagram of a lossy image coding scheme.

Figure 1 shows a simplified block diagram of the components of the encoder in a typical lossy compression system.
It consists mainly of four basic stages. In fact, most systems include a pre-processing stage (P) where, if needed, a color
model conversion or a dimension reduction is performed. Then, firstly, a transform (T) is applied to the input data in
order to obtain decorrelated coefficients and a higher compactness of energy in a few coefficients. Secondly, a quantization
stage (Q) removes information considered unnecessary for user purposes. Thirdly, a bit plane encoding (BPE) is applied
to account for the significance of the quantized coefficients. Fourthly, an entropy coding scheme (EC) is used to reduce
the amount of bits needed to send the significant quantized coefficients through the transmission channel. At the receiver
side, the decoder performs the inverse operations in reverse order. The overall goal is to produce a recovered image as
close as possible to the original image I while preserving the bit rate needed to transmit I ∗ as low as possible, maximizing
a quality criterion. For watermarking purposes, lossy compression is only used to generate a similar image I ∗ , so we will
not discuss any details about the bit plane encoding or entropy coding stages in this paper, which are lossless in nature and
introduce no distortion in the quantized image. Therefore, only preprocessing, transform and quantization are relevant for
determining the final I ∗ .
This general coding scheme must be adapted to the particular characteristics of the source, in order to maximize both
compression ratio and image fidelity. In the case of multispectral and hyperspectral images, the most important issue is
that they are 3D in nature, but this is not exactly true, as two dimensions are spatial but the third one is spectral. An ideal
compression method would take advantage of this fact, trying to exploit both spatial and spectral redundancy. Coding each
band separately is sub-optimal as spectral redundancy is not exploited, while applying 3D coding schemes does not take into
account the difference between spatial and spectral dimensions. Several authors have proposed 3D transformations (usually
DCT or DWT) to decorrelate spatial and spectral redundancy (see 11 for a survey on hyperspectral image compression). In
this paper, we explore the use of vector quantization for lossy compression, as all bands (or groups of bands) are processed
at the same time.

2.1. Vector quantization


Let I be a three dimensional image (or a sub-image block) M × N × d. Each image band is segmented in blocks W × H,
and vectors are created by groping all pixels and spectral values, so they have a length of W × H × d elements. Using
W = 1 and H = 1 leads to spectral vector quantization, as each vector is a spectral signature. Vector quantization 12 is
better from the Shannon’s rate-distortion theory point of view, although it can be computationally prohibitive, so alternative
methods known to be suboptimal need to be explored for practical applications.

Figure 2. 3D blocking for vector quantization.

Regarding the preprocessing stage, we do not apply any other action than the described image blocking. For hyperspec-
tral and ultraspectral images, with a large number of bands, we propose to group bands according to a reordering criterion,
creating different spectral signatures that will be processed separately, following the recommendations in 12 .
In order to exploit both spatial and spectral redundancy, some authors propose to use a DWT first on each band and then
a vector quantization for each new spectral signature (see 13 for example) using a multiresolution approach. We follow the
same framework but we use the original image pixels, without any transformation, because we do not pursue maximizing
compression ratio, but obtaining similar images I ∗ for embedding information for watermarking purposes.
We use a Tree Structured Vector Quantizer (TSVQ), which is known to be suboptimal, but it allows us to manipulate
the final structure of the codebook in order to introduce the appropriate information for watermarking purposes. Basically,
TSVQ works starting with an initial centroid as the codebook, that is, a tree with a single leaf, and then a quality criterion
is applied (mean square error, namely MSE, for example). If there is room for improvement (that is, MSE is greater than
zero), the leaf with higher distortion is split in two, by generating two similar but not equal centroids, and then applying the
Linde-Buzo-Gray algorithm 14 for computing the new centroids. This process is repeated until a general quality criterion
is achieved, or when all leaves contain only equal samples within the same leaf, meaning that a perfect tree T has been
built. For large images, when the number of training vectors is also large, the obtained tree is usually very deep, although
it might be very unbalanced. Nevertheless, the number of possible subtrees is large enough to explore the possibilities of
finding feasible subtrees for watermark embedding.
Then, starting from the final tree T , the optimal subtree in a rate-distortion sense is selected by pruning the tree with the
generalized BFOS algorithm 15 . In this case, distortion is measured by means of MSE, but this criterion could be replaced
with any other more focused on the particular requirements of the watermarking system. The BFOS algorithm ensures the
best possible subtree with the lowest rate for the selected criterion, in the optimal rate-distortion curve generated by all the
optimal subtrees of T .
Finally, the original image is coded using the selected subtree, replacing each original vector by the closest centroid,
that is, by the centroid representing all the elements in the leaf where the original vector falls into starting from the root of
the tree, and selecting the closest centroid until a leaf is reached. Unless generating the initial tree T , coding is a very fast
operation that can be done very efficiently. As a possible improvement of the proposed method, if the maximum distance
between the original vectors and the centroid of the leaf where they fall into is known, any modified version of the centroid
between the original vector and the selected centroid could be used for obtaining a modified image, increasing the capacity
of the watermarking scheme. For simplicity reasons, in this paper we use the computed centroid as the representant for all
the pixels falling in the leaf represented by such centroid.

3. MULTIPLE BAND WATERMARKING SCHEME


The novel watermarking scheme presented in this paper is described in the following sections. In subsection 3.1 the method
of lossy compression and watermarking is described, while subsection 3.2 shows how to locate tampered marked images.
As shown in Section 2, the proposed watermarking scheme for multi and hyperspectral images is based on the TSVQ
lossy coding scheme. Usually, fragile or semifragile watermarking schemes of multiple band images consider only one
band or process each band separately, as those described in 16, 17 . It is possible to work with images using only one band
for computing where and how embed the watermark, but then, when multiband images are marked with the same method,
bands are usually marked separately, or only one band or a subset of bands are marked, which leads to suboptimal solutions.
There are different previous works about satellite image watermarking, such as 18 or19 for example. The first one uses
a satellite image and decomposes it into two mutually orthogonal sub-fields, namely the boundary field and pinned field,
but only uses one band of the satellite image and only one field (pinned) for watermarking purposes. The second paper
presents a semifragile watermarking scheme based on the use of wavelet transforms. The edge and texture of the remote
sensing image are extracted and the watermark is embedded only in the edge character. In this case, once again, only one
image band is marked. We propose to use all the bands at the same time, using a vector quantization approach.

3.1. Mark embedding


The number of available bands in a remote sensing image determines its name, that is, multi, hyper or ultraspectral. In this
paper, to minimize the operations with the images, multispectral images have been chosen to test the scheme, since these
images are smaller. However, this method for semifragile watermarking may be adapted to work as well with the two other
types of remote sensing images. The experiments have been performed for CASI images, which have 14 bands. Due to the
large number of bands of hyperspectral images (224 for an AVIRIS image), we propose to reorder the bands and perform
a partitioned vector quantization, as defined in 20 .
Figure 3 shows three of the fourteen bands of a CASI image from the same area, a volcanic zone of Catalonia, which
comprises a urban area, woods and extinguished volcanoes. Each band shows a different light reflection given by a
special sensor, according to the wavelength assigned to it. Therefore, each pixel of the real image has fourteen spectral
values. These values, which are called the signature of the image, are reorganized as shown in Figure 4. For an image of
Figure 3. Example of multispectral image bands (band 3, band 9 and band 14).

512 × 512 pixels, there are 262.144 signatures with 14 values, one for each band, which is a reasonably large value for
vector quantization purposes.

Figure 4. Generation of the signature vectors for TSVQ from a CASI image.

For detecting the different areas of a possibly tampered marked image with enough detail, the image is divided into
different small regions. This is done following the recommendations in 21 , where it is shown that the ratio between the
number of available samples and vector dimensionality should be at least 30 times, in order to minimize any statistical bias
caused by insufficient number of samples. Therefore, as in our case the dimension of the signatures is 14, we have to have
at least minimum 14 × 30 different samples for compression and classification purposes. For this reason, the regions must
have at least 420 samples and, therefore, the minimum size of these regions is 32 × 32. Nevertheless, as it is described
next, it is better to use regions of 64 × 64 pixels. If necessary, the original image could be further divided in 8 × 8 small
regions, and these 64 small regions would be marked separately with the method introduced below.
The proposed semifragile watermarking method for multispectral images tries to take advantage of the characteristics
of the TSVQ compression method. The information contained in the spectral signatures is used for building the tree which
will be manipulated to satisfy a certain criterion, minimizing the possibility of obtaining an equivalent tree randomly. A
modification in any single band or subset of bands is considered unacceptable, because it introduces an uneven change in
the spectral signature. Only modifications affecting the whole signature are accepted, such as lossy compression, but only
until a certain degree. This makes our proposal different from other semifragile watermarking methods in the literature 16, 17
, which are focused on spatial dimensions rather on the spectral one, processing each band separately.
At this point, is important to note that the embedded mark is somehow determined by the compression rate or, at least,
related to it with a strong correlation. However, the compression rate does not take into account all the individual images
but the signatures as a whole, since we are using a vector quantization approach. This method has been designed explicitly
so that it can work simultaneously with all the bands and the mark in the modified signatures is supposed to preserve the
original information characteristics to the maximum.
First, using the TSVQ algorithm detailed in 12 , a complete or perfect tree of centroids is constructed, containing all
the possible subtrees for compressing of the original image. Then, the BFOS algorithm, called also pruning, prunes the
generated tree with the selected criteria to obtain all the subtrees in the Rate/Distortion curve, namely the convex hull. One
parameter of the resulting TSVQ determines the subtree that is required to obtain the specific compression of the image,
for example the compression rate in itself, distortion or the ratio distortion/compression (namely λ), and selects only one
subtree of the thousands of possible subtrees.
Each of the regions of the original image is compressed with a different compression ratio, for example according to
a known criterion. This new image is passed through the TSVQ again, as shown in 5 using the same parameters as in the
first iteration, and this process is repeated until the generated image has the desired properties.

Figure 5. Embedding block diagram.

To be robust against possible attacks, such as copy-and-replace, or pasting one part of another image into the marked
image, a pseudo-random sequence is chosen to determine the values used as criteria to select the compression subtree in
the pruning algorithm. Thus, it is much more difficult to find a pattern revealing the watermarking scheme properties, thus
reducing the possibilities of manipulating an image or modifying it using another region of the same image, because with
the use of the pseudo-random sequence the contiguous regions do not have similar values of compression rate or any other
used criteria. A seed key is required to watermark images, which is used to generate the pseudo-random sequence. It is
necessary to use as many keys as images are going to be to watermarked with the proposed method, so the key has to be
unique for each multispectral image. This key will be required in the detection process. Nevertheless, the proposed method
does not need the original image for tampering detection, so it can be considered blind.
Finally, the BFOS algorithm generates a table with all the possible subtrees which are in the convex hull, minimizing
distortion for a given compression rate. Usually, both compression rate and MSE are used for generating the convex hull,
but the BFOS algorithm may be used with any other criteria, which might be more suitable for watermarking purposes, or
for joint compression and watermarking 22 . Each value of the pseudo-random sequence is assigned to each region of the
original multispectral image. With these different values, a criterion for each region of the original image is chosen. Each
value represents a different rate compression from each small images.

3.2. Mark detection


The method of detection of the mark for a (possibly modified) image, shown in figure 6, is analogous to the marking
process, and it only requires the key of pseudo-random sequence. The marked image must be divided into smaller regions
(blocks) to determine if they have been modified. Then, with the blocks and the key for that block, the process of mark
detection is performed. The detection of a modification of the image can be performed by checking the criterion for that
sub-image with that key. Therefore, with this method it is possible to inspect and locate tampered marked image.

Figure 6. Detection block diagram.

4. EXPERIMENTAL RESULTS
As described in section 3.1, a multispectral CASI image has been used to evaluate our proposal. The image which we
have used is a part of a big multispectral image of one area of Catalonia, as described above, of 2.485 meters of width
and 15.953 meters of height. Although our proposal can work with any image size (with additional padding if necessary
to obtain complete blocks), we have selected a 512 × 512 image, which represents an area of 1.792 meters of width and
height. The regions in which we have applied the method, of 64 × 64 pixels, represent only 224 × 224 meters of the terrain,
which is a reasonable size for tampering detection purposes. For evaluation purposes, we measure mark influence on image
quality and an example of robustness against copy-and-replace attack is shown.

Table 1. PSNR of marked image for each band.


Band PSNR Band PSNR Band PSNR
1 41.33 6 42.92 11 38.25
2 43.81 7 42.50 12 36.33
3 43.72 8 41.59 13 38.27
4 43.93 9 41.49 14 36.99
5 42.76 10 41.32 Average 41.08

The goal of the scheme is to mark the image by using the spectral signatures. However, in table 1 the PSNR of each
band from the image marked with this method is presented, as it is the common measure for image quality comparisons 8
. The PSNR measures the distortion between two images, and is useful to determine the perturbation of the mark in
watermarking systems for both compression and classification purposes. Notice that PSNR is very high, specially for
central bands containing more information, and it is slightly worse for the last bands, although computed values are still
reasonable.
In Table 2, an example of the different subtrees generated during the pruning algorithm is described. For each region
of the image of 64 × 64 pixels, in which it has been divided, we obtain a table like this one. This table 2 shows, for each
possible subtree that can be generated by pruning the original tree (denoted by index 0), the λ value (i.e., a ratio between
the distortion and the rate compression), the compression rate, the distortion, the maximum and minimum depth of the
subtree, the number of nodes of the subtree and the entropy of the subtree. This measures are used as the basic criteria for
stopping the pruning algorithm and determining the specific subtree that will be used for embed the mark. A simple but
Table 2. Example of computed values during the BFOS algorithm.
subtree λ rate compression distortion num. nodes depth min depth max entropy
0 0.222 13.303 0.0000 6813 6 24 -11.651
1 0.666 13.302 0.0002 6811 6 24 -11.649
2 0.666 13.301 0.0006 6809 6 24 -11.649
3 0.833 13.301 0.0011 6807 6 24 -11.648
4 0.889 13.300 0.0019 6805 6 24 -11.647
5 1.000 13.299 0.0026 6803 6 24 -11.647
6 1.000 13.298 0.0035 6801 6 24 -11.645
7 1.111 13.296 0.0055 6791 6 24 -11.644
8 1.250 13.295 0.0063 6789 6 24 -11.643
... ... ... ... ... ... ... ...
805 10.000 9.5067 19.6232 2867 6 17 -8.640
... ... ... ... ... ... ... ...

efficient criterion is stopping when the maximum depth achieves a desired value, for example. Even with simple criteria
like this, it is possible to generate complex subtrees that are impossible to reproduce by manipulating the original image.
Once it is determined which subtree will be used to generate the modified image, the resulting watermarked image is
constructed with the centroids of the selected subtree which represent the embedded mark. A known key is used to initiate
the pseudo-random sequence, and every value of this sequence is used to choose the subtree for avoiding copy-and-replace
attacks between blocks of the same image.
Figure 7 shows the original image and the marked image with one specific key. Notice that introduced artifacts are
visually undetectable. On the other hand, Figure 8 shows the tampered image and the tamper location detected by the
proposed scheme. In this image it is possible to see a urban region which has been copied and replaced into a forest region,
next to the urban region, which is detected by the suggested scheme, without any other false positive.

Figure 7. Original (left) and marked image (right) for band 9.


Figure 8. Tampered image (left) and tamper location (right).

5. CONCLUSION
In this work we have presented a novel method for multi and hyperspectral semifragile watermarking based on tree struc-
tured vector quantization. In our approach, we use all the information present in all the bands at the same time, thus taking
advantage of both spatial and spectral redundancy for compression purposes. Basically, the original image is segmented
in three dimensional blocks and a tree structured vector quantizer is built for each block. The original block is replaced
by a new one (very similar in a distortion sense) generated by substituting each original vector by the closest centroid in
the selected subtree. This process is repeated until a certain stopping criterion is satisfied. Each block generates a different
subtree and a secret key is used to avoid copy and paste attacks between blocks.
We have also presented an example of mark, attack and detection process with a CASI image, a multispectral image.
Results show that an attack of copying and replacing a region of the image is detected by the scheme. On the other hand,
the mark process achieves, for each band of the image, a high value of PSNR, that is, the marked image is very similar to
the original image and introduced artifacts are visually undetectable. The process of detection is a very simple method, as
a tree is built for each block only once, and the selected tree property is tested. If a block satisfies such property, then it is
not changed, otherwise the detection process marks it as tampered.
Current and future research lines in this subject include the study of optimal subtree selection criteria for pruning
purposes, in order to reduce the region of detection, and, at the same time, to detect the most probable modified pixels
instead of complete regions, increasing tampering detection resolution. A further evaluation of the proposed scheme for
evaluating robustness, capacity and the impact on classification accuracy is under study. Our method needs some tuning
to obtain results similar to those presented in 16 . Finally, although new standards for image compression are becoming
popular, such as the JPEG2000 standard 23 , and these standards use the discrete wavelet transform and process each band
(usually luminance and chrominance bands) separately, it is interesting to explore the possibilities of developing hybrid
systems using TSVQ for for multiple band images, exploiting the extending capabilities of such standard.

ACKNOWLEDGMENTS
and disclaimer
This work is partially supported by the Spanish MCYT and the FEDER funds under grants TIC2003-08604-C04-04 MULTIMARK
and SEG2004-04352-C04-04 PROPRIETAS-WIRELESS. The work described in this paper has been supported in part by the European
Commission through the IST Programme under Contract IST-2002-507932 ECRYPT. The information in this document reflects only
the author’s views, is provided as is and no guarantee or warranty is given that the information is fit for any particular purpose. The user
thereof uses the information at its sole risk and liability.
REFERENCES
1. F. C. Wong and N. Y. Lao, “Economic value of remote sensing imagery for agricultural applications,” in Proceedings
of the Aerospace Conference, 8, pp. 3815–3829, Mar. 2003.
2. P. M. Atkinson and N. J. Tate, Advances in remote sensing and GIS analysis, Wiley, 1999.
3. E. Lin, C. Podilchuk, and E. Delp, “Detection of image alterations using semi-fragile watermarks,” Proceedings of
the SPIE International Conference on Security and Watermarking of Multimedia Contents II 3971, Jan 2000.
4. J. Minguillón, J. Herrera-Joancomartı́, D. Megı́as, and J. Serra-Sagristà, “Evaluation of copyright protection schemes
for hyperspectral imaging,” in Proceedings of Image and signal processing for remote sensing IX, 5238, pp. 512–523,
(Barcelona, Spain), September 2003.
5. NASA, “CASI, compact airborne spectrographic imager.” Available at
http://geo.arc.nasa.gov/sge/jskiles/top-down/OTTER/OTTERdocs/CASI.html.
6. NASA, “AVIRIS, airborne visible/infrared imaging spectrometer.” Available at
http://aviris.jpl.nasa.gov/.
7. J. Serra-Sagristà, C. Fernández, F. Aulı́, F. Garcı́a, and J. Minguillón, “Lossy coding techniques for high-resolution
images,” in Proceedings of Image and signal processing for remote sensing IX, 5238, pp. 276–287, (Barcelona,
Spain), September 2003.
8. E. Cristophe, D. Léger, and C. Mailhes, “Quality criteria benchmark for hyperspectral imagery,” IEEE Transactions
on Geoscience and Remote Sensing 43, pp. 2103–2114, Sept. 2005.
9. J. Minguillón, J. Pujol, J. Serra, and I. Ortuño, “Influence of lossy compression on hyperspectral image classification,”
in Proceedings of Data Mining’2000, pp. 545–554, (Cambridge, UK), July 5-7 2000.
10. J. Minguillón, J. Pujol, J. Serra, I. Ortuño, and P. Guitart, “Adaptive lossy compression and classification of hy-
perspectral images,” in Proceedings of Image and Signal Processing for Remote Sensing VI, 4170, pp. 214–225,
(Barcelona, Spain), Sept. 2000.
11. G. Motta, F. Rizzo, and J. A. Storer, Hyperspectral Data Compression, Springer-Verlag New York, Inc., Secaucus,
NJ, USA, 2005.
12. A. Gersho and R. M. Gray, Vector Quantization and Signal Compression, Communications and Information Theory,
Kluwer Academic Publishers, Norwell, MA, USA, 1992.
13. R. Kumar and V. Makkapati, “Encoding of multispectral and hyperspectral image data using wavelet transform and
gain shape vector quantization.,” Image and Vision Computing 23(8), pp. 721–729, 2005.
14. Y. Linde, A. Buzo, and R. M. Gray, “An algorithm for vector quantizer design,” IEEE Transactions on Communica-
tions COM-28, pp. 84–95, Jan. 1980.
15. L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone, Classification and Regression Trees, Wadsworth Interna-
tional Group, 1984.
16. O. Ekici, B. Sankur, U. Naci, B. Coskun, and M. Akcay, “Comparative assessment of semifragile watermarking
methods,” Journal of Electronic Imaging 13, pp. 209–216, Jan. 2004.
17. C. Rey and J.-L. Dugelay, “A survey of watermarking algorithms for image authentication,” Journal on Applied Signal
Processing 6, pp. 613–621, 2002.
18. A. T. S. Ho, X. Zhu, and W. Woon, “A semi-fragile pinned sine transform watermarking system for content authenti-
cation of satellite images,” IEEE International Geoscience and Remote Sensing Symposium 1-8, 2005.
19. Q. Qin, W. Wang, and S. Chen, “Research of digital semi-fragile watermarking of remote sensing image based on
wavelet analysis,” IEEE International Geoscience and Remote Sensing Symposium 1-7, 2004.
20. B. Huang, A. Ahuja, H. A. Huang, T. J. Schmit, and R. W. Heymann, “Predictive partitioned vector quantization for
hyperspectral sounder data compression,” in Proceedings of the SPIE, 5548, pp. 70–77, Oct. 2004.
21. S. J. Raudys, “On dimensionality, sample size, and classification error of nonparametric linear classification algo-
rithms,” IEEE Transactions on Pattern Analysis and Machine Intelligence 19, pp. 667–671, June 1997.
22. R. Caldelli, G. Macaluso, M. Barni, and E. Magli, “Joint near-lossless watermarking and compression for the au-
thentication of remote sensing images,” in Proceedings of the 24th International Geoscience and Remote Sensing
Symposium, 1, Sept. 2004.
23. D. Santa-Cruz, T. Ebrahimi, J. Askelöf, M. Larsson, and C. Christopoulos, “JPEG 2000 still image coding versus
other standards,” in Proceedings of the SPIE, Applications of Digital Image Processing XXIII, 4115, pp. 446–454,
(San Diego, CA, USA), July 2000.

Vous aimerez peut-être aussi