Gerald J. Dobeck

Algorithm fusion for automated sea mine detection and classification

The fusion of multiple detection/classification algorithms is proving a very powerful approach for dramatically reducing false alarm rate, while still maintaining a high probability of detection and classification. This has been demonstrated in several Navy sea tests. The high-resolution sonar is one of the principal sensors used by the Navy to detect and classify sea mines in mine hunting 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 mine hunting 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). 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. 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).

A Parametric Model for Characterizing Seabed Textures in Synthetic Aperture Sonar Images

High-resolution synthetic aperture sonar (SAS) systems yield finely detailed images of sea bottom environments. SAS image texture models must be capable of representing a wide variety of sea bottom environments including sand ripples, coral or rock formations, and flat hardpack. In this paper, a parameterized model for SAS image textures is derived from the autocorrelation functions (ACFs) of the SAS imaging point spread function (PSF) and the ACF of the seabed texture sonar cross section (SCS). The proposed texture mixture model is analytically tractable and parameterized by component mixing parameters, mixture component correlation lengths, the single-point intensity image statistical shape parameter, and the rotation of the ACF mixture components in the 2-D imaging plane. An iterative parameter estimation algorithm based on the expectation-maximization (EM) algorithm for truncated data is presented and tested against various synthetic and real SAS image textures. The performance of the algorithm is compared and discussed for synthetically generated data across various image sizes and texture characteristics. The model fit is also compared against a small set of real SAS survey images and is shown to accurately fit the imaging PSF and seabed SCS ACF for these textures of interest.

Fusion of adaptive algorithms for the classification of sea mines using high resolution side scan sonar in very shallow water

A new sea mine computer-aided-detection/computer-aided-classification (CAD/CAC) processing string has been developed. The CAD/CAC processing string consists of preprocessing, adaptive clutter filtering (ACF), normalization, detection, feature extraction, classification and fusion processing blocks. The range-dimension ACF is an adaptive linear FIR filter, which is matched both to average highlight and shadow information, while also simultaneously suppressing background clutter. For each detected object, features are extracted and processed through an orthogonalization transformation, enabling an efficient application of the optimal log-likelihood-ratio-test (LLRT) classification rule, in the orthogonal feature space domain. The classified objects of 3 distinct processing strings, developed by 3 different researchers, are fused, using the classification confidence values as features and logic-based, M-out-of-N, or novel LLRT-based fusion rules. The utility of the overall processing strings and their fusion was demonstrated with new very shallow water high-resolution sonar imagery data. The processing string detection and classification parameters were tuned and the string classification performance was optimized, by appropriately selecting a subset of the original feature set. It was shown that LLRT-based fusion algorithms outperform the logic-based or the M-out-of-N ones. The LLRT-based fusion of the CAD/CAC processing strings resulted in a four-fold false alarm rate reduction, compared to the best single CAD/CAC processing string results, while maintaining a constant correct mine classification probability.

A probabilistic model for score-based algorithm fusion

A probabilistic model is described that helps explain how score-based algorithm fusion achieves significant false alarm reduction. The two-class detection and classification (DC) problem is considered: the target and the nontarget. Multiple DC algorithms, based on fundamentally different DC methodologies, process the same sensor data looking for target-like objects. Each object detected and classified as target-like by a given algorithm is assigned a positive score, which indicates the degree to which the algorithm considers the object target-like. Score-based algorithm fusion is the fusion of multiple detection & classification algorithms where only the scores of the individual algorithms are used to make a final determination on whether an object is a target or not Despite the fact that only the scores are used in the fusion process, false alarm reduction has been remarkable while still preserving a high probability of target detection and classification. A probabilistic model is presented in this paper that supports this observation.

The use of texture measures in improving mine classification performance

Research over the last 9 years has resulted in an effective mine classification approach that involves the use of image-segmentation based screening methods followed by multilayer perceptron networks for mine classification. The present approach centers around a baseline 23 Feature set related to highlight, shadow, and highlight/shadow contrast statistic based segmentations, and the use of associated statistical and shape related factors. In the work described here we investigate the improvement of baseline performance by incorporating image texture related features such as cooccurrence matrix related factors.

Kernel-Matching Pursuits With Arbitrary Loss Functions

The purpose of this research is to develop a classifier capable of state-of-the-art performance in both computational efficiency and generalization ability while allowing the algorithm designer to choose arbitrary loss functions as appropriate for a give problem domain. This is critical in applications involving heavily imbalanced, noisy, or non-Gaussian distributed data. To achieve this goal, a kernel-matching pursuit (KMP) framework is formulated where the objective is margin maximization rather than the standard error minimization. This approach enables excellent performance and computational savings in the presence of large, imbalanced training data sets and facilitates the development of two general algorithms. These algorithms support the use of arbitrary loss functions allowing the algorithm designer to control the degree to which outliers are penalized and the manner in which non-Gaussian distributed data is handled. Example loss functions are provided and algorithm performance is illustrated in two groups of experimental results. The first group demonstrates that the proposed algorithms perform equivalent to several state-of-the-art machine learning algorithms on well-published, balanced data. The second group of results illustrates superior performance by the proposed algorithms on imbalanced, non-Gaussian data achieved by employing loss functions appropriate for the data characteristics and problem domain.

Multi-frequency backscatter variations for classification of mine-like targets from low-resolution sonar data

The purpose of this research is to classify bottom targets as mine like using low-resolution sonar data. The sonar is mounted on an autonomous, unmanned seafloor crawler, which operates in the littoral regions. Due to the sonars relatively large beamwidths and proximity to the seafloor, low-resolution imagery is produced in which objects possess Little discernable shape and no acoustic shadow (i.e., image-based classification is infeasible). Therefore, this research forms multiple images with the sonar operating at different, narrowband frequencies. A feature set is then constructed based on backscatter variations across these various narrowband frequency ranges. In addition to representing target response versus frequency, this feature set also possesses the desirable characteristic of range independence. Afterwards, an unsupervised self-organizing map artificial neural network is trained to project and cluster the features into a two-dimensional feature space. In this space, the decision surfaces are formed and the cluster identities are labeled. Experimental results from a variety of target types illustrate the favorable separability of the clusters and the discriminating power of the frequency-based feature set.