244
Maria Calagna and Luigi V. Mancini
sensed satellite images are an important source of this kind of data. They consist of
multispectral images that capture different kinds of information through different
bands, or layers, where each pixel represents a specific position on the Earth and
the pixel brightness value represents the remotely sensed value of some information,
such as water, vegetation, urban areas, and so on.
Usually, features from different bands may be integrated to perform homeland
reasoning and decision making. In general, classification of scientific images, including satellite and medical images, is very important in order to perform high-level
tasks that significantly help field people to elaborate additional views that are created
upon available data.
In [13], the impact of different methods of invisible watermarking against
typical classification algorithms (the migrating means among the iterative optimization algorithms and, the nearest neighbor among the agglomerative hierarchical clustering algorithms) is analyzed. The watermark may be embedded in the spatial domain of the original image, directly or through a transform domain. According to
experimental results, watermark embedding in the spatial domain performs better
than transform domain embedding, since the misclassification percentage is less in
the first case. One possible interpretation concerns the way watermarking distortion is introduced; the first approach tends to distribute changes uniformly, while
the last one tends to distribute changes at those positions that are near the edges in
the image.
11.4 A Proposal for a SVD-based Digital Watermarking Scheme
Within the SPADA@WEB project [27], we developed a blind image watermarking system [5]; only the embedded watermark and the singular values of the cover
image are required in the detection phase, the original cover is not required. This
approach is suitable for Internet applications; in this scenario, the original cover cannot be available to receivers. It works on a block-by-block basis and makes use of
SVD compression to embed the watermark. The watermark is embedded in all the
non-zero singular values according to the local features of the cover image so as to
balance embedding capacity with distortion. The experimental results show that our
scheme is robust, especially with respect to high-pass filtering. To the best of our
knowledge the use of the block-based SVD for watermarking applications has not
yet been investigated.
The following section introduces the basic fundamentals of the SVD transform,
then the SVD-based watermarking algorithm is presented.
11.4.1 The SVD Transform
A common representation of images is in terms of matrices. In the spatial domain,
images are usually represented as matrices of M × N pixels, while in the transform
domain, images are represented as matrices of coefficients. According to SVD, a
matrix may be factorized in three matrices to achieve the goal of compression [15].
Given a generic matrix A with M rows and N columns, A
M×N (M ≤ N), then the
Maria Calagna and Luigi V. Mancini
sensed satellite images are an important source of this kind of data. They consist of
multispectral images that capture different kinds of information through different
bands, or layers, where each pixel represents a specific position on the Earth and
the pixel brightness value represents the remotely sensed value of some information,
such as water, vegetation, urban areas, and so on.
Usually, features from different bands may be integrated to perform homeland
reasoning and decision making. In general, classification of scientific images, including satellite and medical images, is very important in order to perform high-level
tasks that significantly help field people to elaborate additional views that are created
upon available data.
In [13], the impact of different methods of invisible watermarking against
typical classification algorithms (the migrating means among the iterative optimization algorithms and, the nearest neighbor among the agglomerative hierarchical clustering algorithms) is analyzed. The watermark may be embedded in the spatial domain of the original image, directly or through a transform domain. According to
experimental results, watermark embedding in the spatial domain performs better
than transform domain embedding, since the misclassification percentage is less in
the first case. One possible interpretation concerns the way watermarking distortion is introduced; the first approach tends to distribute changes uniformly, while
the last one tends to distribute changes at those positions that are near the edges in
the image.
11.4 A Proposal for a SVD-based Digital Watermarking Scheme
Within the SPADA@WEB project [27], we developed a blind image watermarking system [5]; only the embedded watermark and the singular values of the cover
image are required in the detection phase, the original cover is not required. This
approach is suitable for Internet applications; in this scenario, the original cover cannot be available to receivers. It works on a block-by-block basis and makes use of
SVD compression to embed the watermark. The watermark is embedded in all the
non-zero singular values according to the local features of the cover image so as to
balance embedding capacity with distortion. The experimental results show that our
scheme is robust, especially with respect to high-pass filtering. To the best of our
knowledge the use of the block-based SVD for watermarking applications has not
yet been investigated.
The following section introduces the basic fundamentals of the SVD transform,
then the SVD-based watermarking algorithm is presented.
11.4.1 The SVD Transform
A common representation of images is in terms of matrices. In the spatial domain,
images are usually represented as matrices of M × N pixels, while in the transform
domain, images are represented as matrices of coefficients. According to SVD, a
matrix may be factorized in three matrices to achieve the goal of compression [15].
Given a generic matrix A with M rows and N columns, A
M×N (M ≤ N), then the