Here we discuss an open software architecture to develop, test, and evaluate machine learning algorithms for target detection and classification. This architecture, known as the Modular Algorithm Testbed Suite (MATS), aids developers by defining interfaces for various portions of the automatic target detection processing chain. There are several key advantages to this approach. First, with “plug and play” modules for detection, feature extraction, and classification, developers can mix and match different approaches and focus on particular portions of the processing chain that yield the most performance benefit. Second, since some portions of the processing chain may be more agnostic to the sensor data type than others, e.g. target features may change but the pattern classification approach is the same, MATS enables quick ATR development for similar data types. Finally, since developers can insert "black boxes" into the ATR processing chain, MATS allows for independent blind testing of algorithms without compromising intellectual property. In this paper, we will discuss the MATS architecture and review several case studies where MATS enabled rapid demonstration and transition of ATR algorithms to Navy mine countermeasure (MCM) post-mission analysis software.
Superpixel segmentation methods have been found to be increasingly valuable in image processing and analysis. Superpixel segmentation approaches have been used as a preprocessing step for a wide variety of image analysis tasks such as full scene segmentation, automated scene understanding, object detection and classification, and have been used to reduce computation time during these tasks. While many quantitative evaluation metrics have been developed in the literature to analyze traditional image segmentation and clustering results, these metrics have not been used or adapted to quantitatively evaluate superpixel segmentations. In this paper, multiple superpixel segmentation algorithms are applied to synthetic aperture sonar (SAS) imagery and the results are evaluated using cluster validity indices that have been adapted for superpixel segmentation. Both cluster validity metrics that rely only on internal measures as well as those that use both internal and external measures are considered. Results are shown on a synthetic aperture sonar (SAS) data set.
An approach to image labeling by seabed context based on multiple-instance learning via embedded instance selection (MILES) is presented. Sonar images are first segmented into superpixels with associated intensity and texture feature distributions. These superpixels are defined as the "instances" and the sonar images are defined as the "bags" within the MILES classification framework. The intensity feature distributions are discrete while the texture feature distributions are continuous, thus the Cauchy-Schwarz divergence metric is used to embed the instances in a higher-dimensional discriminatory space. Results are given for labeled synthetic aperture sonar (SAS) image database containing images with a variety of seabed textures.
Characteristics of underwater targets displayed in synthetic aperture sonar (SAS) imagery vary depending on their environmental context. Discriminative features in sea grass may differ from the features that are discriminative in sand ripple, for example. Environmentally-adaptive target detection and classification systems that take into account environmental context, therefore, have the potential for improved results. This paper presents an end-to-end environmentally-adaptive target detection system for SAS imagery that performs target recognition while accounting for environmental context. First, locations of interest are identified in the imagery using the Reed-Xiaoli (RX) detector and a Non-Gaussian detector based on the multivariate Laplace distribution. Then, the Multiple Instance Learning via Embedded Instance Selection (MILES) approach is used to identify the environmental context of the targets. Finally, target features are extracted and a set of environmentally-specific k-Nearest Neighbors (k-NN) classifiers are applied. Experiments were conducted on a collection of both high and low frequency SAS imagery with a variety of environmental contexts and results show improved classification accuracy between target classes when compared with classification results with no environmental consideration.
There is a desire in the Mine Counter Measure community to develop a systematic method to predict and/or estimate the performance of Automatic Target Recognition (ATR) algorithms that are detecting and classifying mine-like objects within sonar data. Ideally, parameters exist that can be measured directly from the sonar data that correlate with ATR performance. In this effort, two metrics were analyzed for their predictive potential using high frequency synthetic aperture sonar (SAS) images. The first parameter is a measure of contrast. It is essentially the variance in pixel intensity over a fixed partition of relatively small size. An analysis was performed to determine the optimum block size for this contrast calculation. These blocks were then overlapped in the horizontal and vertical direction over the entire image. The second parameter is the one-dimensional K-shape parameter. The K-distribution is commonly used to describe sonar backscatter return from range cells that contain a finite number of scatterers. An Ada-Boosted Decision Tree classifier was used to calculate the probability of classification (Pc) and false alarm rate (FAR) for several types of targets in SAS images from three different data sets. ROC curves as a function of the measured parameters were generated and the correlation between the measured parameters in the vicinity of each of the contacts and the ATR performance was investigated. The contrast and K-shape parameters were considered separately. Additionally, the contrast and K-shape parameter were associated with background texture types using previously labeled high frequency SAS images.
This paper proposes a possibilistic context identification approach for synthetic aperture sonar (SAS) imagery. SAS seabed imagery can display a variety of textures that can be used to identify seabed types such as sea grass, sand ripple and hard-packed sand, etc. Target objects in SAS imagery often have varying characteristics and features due to changing environmental context. Therefore, methods that can identify the seabed environment can be used to assist in target classification and detection in an environmentally adaptive or context-dependent approach. In this paper, a possibilistic context identification approach is used to identify the seabed contexts. Alternative methods, such as crisp, fuzzy or probabilistic methods, would force one type of context on every sample in the imagery, ignoring the possibility that the test imagery may include an environmental context that has not yet appeared in the training process. The proposed possibilistic approach has an advantage in that it can both identify known contexts as well as identify when an unknown context has been encountered. Experiments are conducted on a collection of SAS imagery that display a variety of environmental features.
In this paper we describe an unsupervised approach to seabed co-segmentation over the multiple sonar images collected in sonar surveys. We adapt a traditional single image segmentation texton-based approach to the sonar survey task by modifying the texture extraction filter bank to better model possible sonar image textures. Two different algorithms for building a universal texton library are presented that produce common pixel labels across multiple images. Following pixel labeling with the universal texton library, images are quantized into superpixels and co-segmented using a DP clustering algorithm. The segmentation results for both texton library selection criteria are contrasted and compared for a labeled set of SAS images with various discernable textures.
A synthetic aperture sonar (SAS) image segmentation algorithm using features from a parameterized intensity
image autocorrelation function (ACF) is presented. A modification over previous parameterized ACF models that
better characterizes periodic or rippled seabed textures is presented and discussed. An unsupervised multiclass
k-means segmentation algorithm is proposed and tested against a set of labeled SAS images. Segmentation
results using the various models are compared against sand, rock, and rippled seabed environments.
Our previous work developed an online learning Bayesian framework (dynamic tree) for data organization and
clustering. To continuously adapt the system during operation, we concurrently seek to perform outlier detection
to prevent them from incorrectly modifying the system. We propose a new Bayesian surprise metric to differentiate
outliers from the training data and thus help to selectively adapt the model parameters. The metric is
calculated based on the difference between the prior and the posterior distributions on the model when a new
sample is introduced. A good training datum would sufficiently but not excessively change the model; consequently,
the difference between the prior and the posterior distributions would be reasonable to the amount of
new information present on the datum. However, an outlier carries an element of surprise that would significantly
change the model. In such a case, the posterior distribution would greatly differ from the prior resulting in a large
value for the surprise metric. We categorize this datum as an outlier and other means (e.g. human operator) will
have to be used to handle such cases. The surprise metric is calculated based on the model distribution, and as
such, it adapts with the model. The surprise factor is dependent on the state of the system. This speeds up the
learning process by considering only the relevant new data. Both the model parameters and even the structure
of the dynamic tree can be updated under this approach.
KEYWORDS: Data modeling, Data fusion, Statistical modeling, Visual process modeling, Annealing, Algorithms, Statistical inference, Machine vision, Computer vision technology, Chemical elements
In this work, we propose DEformable BAyesian Networks (DEBAN), a probabilistic graphical model framework
where model selection and statistical inference can be viewed as two key ingredients in the same iterative
process. While this concept has shown successful results in computer vision community,1-4 our proposed approach
generalizes the concept such that it is applicable to any data type. Our goal is to infer the optimal structure/model
to fit the given observations. The optimal structure conveys an automatic way to find not only the number of
clusters in the data set, but also the multiscale graph structure illustrating the dependence relationship among
the variables in the network. Finally, the marginal posterior distribution at each root node is regarded as the
fused information of its corresponding observations, and the most probable state can be found from the maximum
a posteriori (MAP) solution with the uncertainty of the estimate in the form of a probability distribution which
is desired for a variety of applications.
An approach to simulate synthetic aperture sonar (SAS) images with known autocorrelation functions (ACF)
and single-point statistics is presented. ACF models for generating textures with and without periodicities are
defined and explained. Simulated textures of these models are compared visually with real SAS image textures.
Distortion and degradation of the synthetic textures are examined for various simulation parameter choices.
This paper describes our attempts to model sea bottom textures in high-frequency synthetic aperture sonar imagery using
a Gaussian Markov random field. A least-squares estimation technique is first used to estimate the model parameters of
the down-sampled grey-scale sonar images. To qualitatively measure estimation results, a fast sampling algorithm is then
used to synthesize the sea bottom textures of a fourth-order Gaussian Markov random field which is then compared with
the original sonar image. A total of four types of sea floor texture are used in the case study. Results show that the 4th
order GMRF model mimics patchy sandy textures and sand ripple, but does not reproduce more complex textures
exhibited by coral and rock formations.
Single-point statistical properties of envelope-detected data such as signal returns from synthetic aperture radar
and sonar have traditionally been modeled via the Rayleigh distribution and more recently by the K-distribution.
Two-dimensional correlations that occur in textured non-Gaussian imagery are more difficult to model and
estimate than Gaussian textures due to the nonlinear transformations of the time series data that occur during
envelope detection. In this research, textured sonar imagery is modeled by a correlated K-distribution. The
correlated K-distribution is explained via the compound representation of the one-dimensional K-distribution
probability density function. After demonstrating the model utility using synthetically generated imagery, model
parameters are estimated from a set of textured sonar images using a nonlinear least-squares fit algorithm. Results
are discussed with regard to texture segmentation applications.
High-resolution sonar images of the sea floor contain rich spatial information that varies widely depending on
survey location, sea state, and sensor platform-induced artifacts. Automatically segmenting sonar images into
labeled regions can have several useful applications such as creating high-resolution bottom maps and adapting
automatic target recognition schemes to perform optimally given the measured environment. This paper presents
a method for sonar image segmentation using graphical models known as dynamic trees (DTs). A DT is a mixture
of simply-connected tree-structured Bayesian networks (TSBNs), a hierarchical two-dimensional Bayesian
network, where the leaf node states of each TSBN are the label of each image pixel. The DT segmentation
task is to find the best TSBN mixture that represents the underlying data. A novel use of the K-distribution
as a likelihood function for associating sonar image pixels with the appropriate bottom-type label is introduced.
A simulated annealing stochastic search method is used to determine the maximum a posteriori (MAP) DT
quadtree structure for each sonar image. Segmentation results from several images are presented and discussed.
KEYWORDS: LIDAR, Data modeling, Motion measurement, Data centers, Data analysis, Motion models, Error analysis, Laser scanners, 3D scanning, Vegetation
Land surface elevation measurements from airborne laser swath mapping (ALSM) data can be irregularly spaced due to occlusion by forest canopy or scanner and aircraft motion. The measurements are usually interpolated into a regularly spaced grid using techniques such as Kriging or spline-interpolation. In this paper a probabilistic
graphical model called a Bayesian network (BN) is employed to interpolate missing data. A grid of nodes is imposed over ALSM measurements and the elevation information at each node is estimated using two methods: 1) a simple causal method, similar to a Markov mesh random field (MMRF), and 2) BN belief propagation. The interpolated results of both algorithms using the maximum a posteriori (MAP) estimates are presented and compared. Finally, uncertainty measures are introduced and evaluated against the final estimates from the BN belief propagation algorithm.
Automatic detection of sea mines in coastal regions is a difficult task due to the highly variable sea bottom conditions present in the underwater environment. Detection systems must be able to discriminate objects which vary in size, shape, and orientation from naturally occurring and man-made clutter. Additionally, these automated systems must be computationally efficient to be incorporated into unmanned underwater vehicle (UUV) sensor systems characterized by high sensor data rates and limited processing abilities. Using noncommutative group harmonic analysis, a fast, robust sea mine detection system is created. A family of unitary image transforms associated to noncommutative groups is generated and applied to side scan sonar image files supplied by Naval Surface Warfare Center Panama City (NSWC PC). These transforms project key image features, geometrically defined structures with orientations, and localized spectral information into distinct orthogonal components or feature subspaces of the image. The performance of the detection system is compared against the performance of an independent detection system in terms of probability of detection (Pd) and probability of false alarm (Pfa).
The high-resolution sonar is one of the principal sensors used by the Navy to detect and classify sea mines in minehunting operations. For such sonar systems, substantial effort has been devoted to the development of automated detection and classification (D/C) algorithms. These have been spurred by several factors including (1) aids for operators to reduce work overload, (2) more optimal use of all available data, and (3) the introduction of unmanned minehunting systems. The environments where sea mines are typically laid (harbor areas, shipping lanes, and the littorals) give rise to many false alarms caused by natural, biologic, and man-made clutter. The objective of the automated D/C algorithms is to eliminate most of these false alarms while still maintaining a very high probability of mine detection and classification (PdPc). In recent years, the benefits of fusing the outputs of multiple D/C algorithms have been studied. We refer to this as Algorithm Fusion. The results have been remarkable, including reliable robustness to new environments. This paper describes a method for training several multivariate Gaussian classifiers such that their And-ing dramatically reduces false alarms while maintaining a high probability of classification. This training approach is referred to as the Focused- Training method. This work extends our 2001-2002 work where the Focused-Training method was used with three other types of classifiers: the Attractor-based K-Nearest Neighbor Neural Network (a type of radial-basis, probabilistic neural network), the Optimal Discrimination Filter Classifier (based linear discrimination theory), and the Quadratic Penalty Function Support Vector Machine (QPFSVM). Although our experience has been gained in the area of sea mine detection and classification, the principles described herein are general and can be applied to a wide range of pattern recognition and automatic target recognition (ATR) problems.
The high-resolution sonar is one of the principal sensors used by the Navy to detect and classify sea mines in minehunting operations. For such sonar systems, substantial effort has been devoted to the development of automated detection and classification (D/C) algorithms. These have been spurred by several factors including (1) aids for operators to reduce work overload, (2) more optimal use of all available data, and (3) the introduction of unmanned minehunting systems. The environments where sea mines are typically laid (harbor areas, shipping lanes, and the littorals) give rise to many false alarms caused by natural, biologic, and man-made clutter. The objective of the automated D/C algorithms is to eliminate most of these false alarms while still maintaining a very high probability of mine detection and classification (PdPc). In recent years, the benefits of fusing the outputs of multiple D/C algorithms have been studied. We refer to this as Algorithm Fusion. The results have been remarkable, including reliable robustness to new environments. The Quadratic Penalty Function Support Vector Machine (QPFSVM) algorithm to aid in the automated detection and classification of sea mines is introduced in this paper. The QPFSVM algorithm is easy to train, simple to implement, and robust to feature space dimension. Outputs of successive SVM algorithms are cascaded in stages (fused) to improve the Probability of Classification (Pc) and reduce the number of false alarms. Even though our experience has been gained in the area of sea mine detection and classification, the principles described herein are general and can be applied to fusion of any D/C problem (e.g., automated medical diagnosis or automatic target recognition for ballistic missile defense).
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.