In this paper, we consider detecting man-made objects in natural images. We segment the image into tiles; we
consider a variety of statistical metrics and correlate them to the presence of man-made targets. To quantify
the metric, we apply a method of implanting targets and evaluating the resulting ROC (Receiver Operating
Characteristic) curves. We rank previously reported algorithms and develop new ones in this paper.
Spectral image complexity is an ill-defined term that has been addressed previously in terms of dimensionality, multivariate normality, and other approaches. Here, we apply the concept of the linear mixture model to the question of spectral image complexity at spatially local scales. Essentially, the "complexity" of an image region is related to the volume of a convex set enclosing the data in the spectral space. The volume is estimated as a function of increasing dimensionality (through the use of a set of endmembers describing the data cloud) using the Gram Matrix approach. It is hypothesized that more complex regions of the image are composed of multiple, diverse materials and will thus occupy a larger volume in the hyperspace. The ultimate application here is large area image search without a priori information regarding the target signature. Instead, image cues will be provided based on local, relative estimates of the image complexity. The technique used to estimate the spectral image complexity is described and results are shown for representative image chips and a large area flightline of reflective hyperspectral imagery. The extension to the problem of large area search will then be described and results are shown for a 4-band multispectral image.
Liquid crystal tunable filters (LCTFs) are a technology that can act as both a spectral and linear polarization filter for an imaging device. Paired with the appropriate hardware, a LCTF can be configured to collect hyperspectral Stokes imagery which contains both spectral as well as polarimetric information on a per-pixel level basis. This data is used to investigate the utility of spectro-polarimetric data with standard spectral analysis algorithms, in this case anomaly detection. A method to simulate different ground sample distances (GSDs) is used to illustrate the effect on algorithm performance. In this paper, a spectro-polarimetric imager is presented that can collect spectro-polarimetric image cubes in units of calibrated sensor reaching radiance. The system is used to collect imagery of two scenes, each containing die-cast scale vehicles and different background types. An anomaly detector is applied to the intensity and polarized image cubes to find those pixels that are different from the background spectrally and/or polarimetrically. The effect of changing the apparent GSD on the anomaly detection performance is explored. This shows that applying anomaly detection to spectro-polarimetric data can improve the false alarm rate over standard spectral data for finding certain types of man-made objects in complex backgrounds.
Many spectral algorithms that are routinely applied to spectral imagery are based on the following models:
statistical, linear mixture, and linear subspace. As a result, assumptions are made about the underlying distribution
of the data such as multivariate normality or other geometric restrictions. Here we present a graph based
model for spectral data that avoids these restrictive assumptions and apply graph based metrics to quantify
certain aspects of the resulting graph. The construction of the spectral graph begins by connecting each pixel to
its k-nearest neighbors with an undirected weighted edge. The weight of each edge corresponds to the spectral
Euclidean distance between the adjacent pixels. The number of nearest neighbors, k, is chosen such that the
graph is connected i.e., there is a path from each pixel xi to every other. This requirement ensures the existence
of inter-cluster connections which will prove vital for our application to change detection. Once the graph
is constructed, we calculate a metric called the Normalized Edge Volume (NEV) that describes the internal
structural volume based on the vertex connectivity and weighted edges of the graph. Finally, we demonstrate
a graph based change detection method that applies this metric.
Historically, much of spectral image analysis revolves around assumptions of multivariate normality. If the background
spectral distribution can be assumed to be multivariate normal, then algorithms for anomaly detection,
target detection, and classification can be developed around that assumption. However, as the current generation
sensors typically have higher spatial and/or spectral resolution, the spectral distribution complexity of the data
collected is increasing and these assumptions are no longer adequate, particularly image-wide. However, large
portions of the imagery may be accurately described by a multivariate normal distribution. A new empirical
method for assessing the multivariate normality of a hyperspectral distribution is presented here. This method
assesses the multivariate normality of individual spectral image tiles and is applied to the large area search problem.
Additionally, the methodology is applied to a selection of full hyperspectral data sets for general content
evaluation. This information can be used to indicate the degree of multivariate normality (or complexity) of the
data or data regions and to determine the appropriate algorithm to use globally or locally for spatially adaptive
processing.
Linear spectral unmixing and endmember selection are two of the many tasks that can be accomplished using hyperspectral
imagery. The quality of the unmixing results depends on an accurate estimate of the number of endmembers used in
the analysis. Too many estimated endmembers produce over fitting of the spectral unmixing results; too few estimated
endmembers produce spectral unmixing results with large residual errors. Several statistical and geometrical approaches
have been developed to estimate the number of endmembers, but many of these approaches rely on using the global
dataset. The global approach does not take into consideration local endmember variability, which is of particular interest
in high-spatial resolution imagery. Here, the number of endmembers within local image tiles is estimated by using a novel,
spatially adaptive approach. Each pixel is unmixed using the locally identified endmembers and global abundance maps
are generated by clustering these locally derived endmembers. Comparisons are made between this new approach and
an established global method that uses PCA to estimate the number of endmembers and SMACC to identify the spectra.
Multiple images with varying spatial resolution are used in the comparison of methodologies and conclusions are drawn
based on per-pixel residual unmixing errors.
Visualization of the high-dimensional data set that makes up hyperspectral images necessitates a dimensionality
reduction approach to make that data useful to a human analyst. The expression of spectral data as color images,
individual pixel spectra plots, principal component images, and 2D/3D scatter plots of a subset of the data are
a few examples of common techniques. However, these approaches leave the user with little ability to intuit
knowledge of the full N-dimensional spectral data space or to directly or easily interact with that data. In this
work, we look at developing an interactive, intuitive visualization and analysis tool based on using a Poincaré
disk as a window into that high dimensional space. The Poincaré disk represents an infinite, two-dimensional
hyperbolic space such that distances and areas increase exponentially as you move farther from the center of the
disk. By projecting N-dimensional data into this space using a non-linear, yet relative distance metric preserving
projection (such as the Sammon projection), we can simultaneously view the entire data set while maintaining
natural clustering and spacing. The disk also provides a means to interact with the data; the user is presented
with a "fish-eye" view of the space which can be navigated and manipulated with a mouse to "zoom" into clusters
of data and to select spectral data points. By coupling this interaction with a synchronous view of the data as
a spatial RGB image and the ability to examine individual pixel spectra, the user has full control over the data
set for classification, analysis, and instructive use.
Change detection with application to wide-area search seeks to identify where interesting activity has occurred
between two images. Since there are many different classes of change, one metric may miss a particular type of
change. Therefore, it is potentially beneficial to select metrics with complementary properties. With this idea
in mind, a new change detection scheme was created using mean-shift and outlier-distance metrics. Using these
metrics in combination should identify and characterize change more completely than either individually. An
algorithm using both metrics was developed and tested using registered sets of multispectral imagery.
In the task of automated anomaly detection, it is desirable to find regions within imagery that contain man-made structures
or objects. The task of separating these signatures from the scene background and other naturally occurring anomalies
can be challenging. This task is even more difficult when the spectral signatures of the man-made objects are designed to
closely match the surrounding background. As new sensors emerge that can image both spectrally and polarimetrically, it
is possible to utilize the polarimetric signature to discriminate between many types of man-made and natural anomalies.
One type of passive imaging system that allows for spetro-polarimetric data to be collected is the pairing of a liquid crystal
tunable filter (LCTF) with a CCD camera thus creating a spectro-polarimetic imager (SPI). In this paper, an anomaly
detection scheme is implemented which makes use of the spectral Stokes imagery collected by this sensing system. The
ability for the anomaly detector to find man-made objects is assessed as a function of the number of spectral bands available
and it is shown that low false alarm rates can be achieved with relatively few spectral bands.
A new method for change detection of two large area scenes based on the point density of the pixel distribution in
the hyperspace is presented. This method is derived from the point density approach to hyperspectral analysis,
originally developed for material discrimination based on inherent dimension estimation. In this method, two
registered large area scenes are tiled for individual scoring and comparison. The point density tail length
is estimated for each tile in both scenes. The difference between this value for corresponding tiles indicates
whether change has likely occurred in a tile and how significant the change is relative to other changes in the
image. The method does not identify changes in individual pixels, but uses a tiling approach to identify changes
in small sub-regions of the image. Preliminary results of this methodology are presented for multiple images and
changing scene phenomenology.
The inherent dimension of hyperspectral data is commonly estimated for the purpose of dimension reduction.
However, the dimension estimate itself may be a useful measure for extracting information about hyperspectral
data, including scene content, complexity, and clutter. There are many ways to estimate the inherent dimension
of data, each measuring the data in a different way. This paper compares a group of dimension estimation metrics
on a variety of data, both full scene and individual material regions, to determine the relationship between the
different estimates and what features each method is measuring when applied to complex data.
The inherent dimension of hyperspectral data may be a useful metric for discriminating between the presence of
manmade and natural materials in a scene without reliance on spectral signatures take from libraries. Previously,
a simple geometric method for approximating the inherent dimension was introduced along with results from
application to single material clusters. This method uses an estimate of the slope from a graph based on the
point density estimation in the spectral space. Other information can be gathered from the plot which may aid
in the discrimination between manmade and natural materials. In order to use these measures to differentiate
between the two material types, the effect of the inclusion of manmade pixels on the phenomenology of the
background distribution must be evaluated. Here, a procedure for injecting manmade pixels into a natural
region of a scene is discussed. The results of dimension estimation on natural scenes with varying amounts of
manmade pixels injected are presented here, indicating that these metrics can be sensitive to the presence of
manmade phenomenology in an image.
The inherent dimensionality of a spectral image can be estimated in a number of ways, primarily based on statistical measures of the data cloud in the hyperspace. Methods using the eigenvalues from Principal Components Analysis, a Minimum Noise Fraction transformation, or the Virtual Dimensionality algorithm are widely applied to entire hyperspectral images typically with the goal of reducing the overall dimensionality of an image. Additionally, in complex scenes containing non-natural materials, the lack of multivariate normality of the data set implies that a statistically-based estimation is less than optimal. However, it is desirable to understand the dimensionality of individual components and small multi-material clusters within a hyperspectral scene, as there is no a priori reason to expect all distinct material classes in the scene to have the same inherent dimensionality. For this reason, the inherent dimensionality may be useful as an indicator of the presense of manmade or natural materials within small subsets of the image. Here, a geometric approach is developed based on the spatially local estimation of dimensionality in the native data hyperspace. It will be shown that the dimensionality of a collection of data points in the full n dimensions (where n is the number of spectral channels measured) can be estimated by calculating the change in point density as a function of distance in the full n dimensional hyperspace. Simple simulated examples to demonstrate the concept will be shown, as well as applications to real hyperspectral imagery collected with the HyMAP sensor.
The inherent dimensionality of a spectral image can be estimated in a number of ways, primarily based on statistical
measures of the data cloud in the hyperspace. Methods using the eigenvalues from a Principal Components
Analysis, a Minimum Noise Fraction transformation, or the Virtual Dimensionality algorithm are widely used
as applied to entire images typically with the goal of reducing the dimensionality of an image in its entirety.
However, it is desirable to understand the dimensionality of individual components within a hyperspectral scene,
as there is no a priori reason to expect all distinct material classes in the scene to have the same inherent dimensionality.
Additionally, in complex scenes containing non-natural materials, the lack of multivariate normality
of the data set implies that a statistically based estimation is less than optimal. Here, a geometric approach is
developed based on the local estimation of dimensionality in the native data hyperspace. It will be shown that
the dimensionality of a collection of data points (k) in the full n dimensions (where n is the number of spectral
channels measured) can be estimated by calculating the change in point density as a function of distance in the
full n dimensional hyperspace. Simple simulated examples to demonstrate the concept will be shown, as well as
applications to real hyperspectral imagery collected with the HyMAP sensor.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.