Frances B. Shin

Modeling chaotic systems with hidden Markov models

The problem of modeling chaotic nonlinear dynamical systems using hidden Markov models is considered. A hidden Markov model for a class of chaotic systems is developed from noise-free observations of the output of that system. A combination of vector quantization and the Baum-Welch algorithm is used for training. The importance of this combined iterative approach is demonstrated. The model is then used for signal separation and signal detection problems. The difference between maximum likelihood signal estimation and maximum a posteriori signal estimation using a hidden Markov model is illustrated for a nonlinear dynamical system.<<ETX>>

Active impulsive echo discrimination in shallow water by mapping target physics-derived features to classifiers

One of the most difficult challenges in shallow-water active sonar processing is false-alarm rate reduction via active classification. In impulsive-echo-range processing, an additional challenge is dealing with stochastic impulsive source variability. The goal of active classification is to remove as much clutter as possible while maintaining an acceptable detection performance. Clutter in this context refers to any non-target, threshold-crossing cluster event. In this paper, we present a clutter-reduction algorithm using an integrated pattern-recognition paradigm that spans a wide spectrum of signal and image processing-target physics, exploration of projection spaces, feature optimization, and mapping the decision architecture to the underlying good-feature distribution. This approach is analogous to a classify-before-detect strategy that utilizes multiple informations to arrive at the detection decision. After a thorough algorithm evaluation with real active sonar data, we achieved over an order of magnitude performance improvement in clutter reduction with our methodology over that of the baseline processing.

Automatic road-distress classification and identification using a combination of hierarchical classifiers and expert systems-subimage and object processing

Automatic recognition of various road distresses is of considerable interest since it facilitates preventive road maintenance before cracks and potholes become too severe, leading to economic benefits. The current approach of using human operators to categorize road distresses is both labor-intensive and time consuming. We describe a two-step algorithm that automates road-distress identification with high accuracy. After constant-false-alarm-rate (CFAR) detection at the pixel level, subimage processing classifies each subimage of 64/spl times/64 pixels (each pixel is 1 mm by 1 mm) into crack, patch/pothole (P2), sealed crack, and false alarm. Object processing performs spatial clustering and object segmentation prior to final distress identification. The major challenge is integrating a number of signal and image processing algorithms to effectively deal with false alarms, film artifacts, and nonstationary distress characteristics and background. We explore how various signal and image processing concepts in signal projection, nonlinear filtering, feature optimization, image coding, and pattern recognition can be judiciously combined for computationally efficient and robust identification of road distresses. Our data analysis of 112 image frames (each frame contains 6144/spl times/4095 pixels) shows that the overall system performance at the object level is as follows: a P/sub D/ of 0.90 (average of 74 subimages per detected object), probability of correct distress identification of 0.96, and a P/sub FA/ of 0.79 false objects (average of 11 subimages per false object) per image frame.

IER clutter reduction in shallow water

One of the most difficult challenges in shallow water active sonar processing is false alarm rate reduction via active classification. The goal of active classification is to remove as much clutter as possible while maintaining an acceptable P/sub d/ performance. Clutter in this context refers to any nontarget, threshold crossing cluster events. We present a novel approach to clutter reduction in shallow water using an integrated pattern recognition paradigm. With our understanding of target physics, we identify clues or features useful for clutter rejection by projecting raw data onto appropriate transformation spaces to achieve both energy compaction and subspace filtering. We explore various time-frequency distribution (TFD) functions, compressed phase map, and speech related processing domains to extract clues that provide insights into local multipath patterns and detailed signal dynamics. We perform thorough feature fusion and optimization, and evaluate clutter reduction performance in terms of receiver operating characteristic (ROC) curves. Our real data analyses indicate that we can achieve over an order of magnitude performance improvement in clutter reduction over that of the baseline processing.

Integrated approach to bandwidth reduction and mine detection in shallow water with reduced-dimension image compression and automatic target recognition algorithms

In distributed underwater signal processing for area surveillance and sanitization during regional conflicts, it is often necessary to transmit raw imagery data to a remote processing station for detection-report confirmation and more sophisticated automatic target recognition (ATR) processing. Because of he limited bandwidth available for transmission, image compression is of paramount importance. At the same time, preservation of useful information that contains essential signal attributes is crucial for effective mine detection and classification in shallow water. In this paper, we present an integrated processing strategy that combines image compression and ATR algorithms for superior detection performance while achieving maximal bandwidth reduction. Our reduced-dimension image compression algorithm comprises image-content classification for the subimage-specific transformation, principal component analysis for further dimension reduction, and vector quantization to obtain minimal information state. Next, using an integrated pattern recognition paradigm, our ATR algorithm optimally combines low-dimensional features and an appropriate classifier topology to extract maximum recognition performance from reconstructed images. Instead of assessing performance of the image compression algorithm in terms of commonly used peak signal-to-noise ratio or normalized mean-squared error criteria, we quantify our algorithm performance using a metric that reflects human and operational factors - ATR performance. Our preliminary analysis based on high-frequency sonar real data indicates that we can achieve a compression ratio of up to 57:1 with minimal sacrifice in PD and PFA. Furthermore, we discuss the concept of the classification Cramer-Rao bound in terms of data compression, sufficient statistics, and class separability to quantify the extent to which a classifier approximates the Bayes classifier.

Bandwidth reduction of high-frequency sonar imagery in shallow water using content-adaptive hybrid image coding

One of the biggest challenges in distributed underwater mine warfare for area sanitization and safe power projection during regional conflicts is transmission of compressed raw imagery data to a central processing station via a limited bandwidth channel while preserving crucial target information for further detection and automatic target recognition processing. Moreover, operating in an extremely shallow water with fluctuating channels and numerous interfering sources makes it imperative that image compression algorithms effectively deal with background nonstationarity within an image as well as content variation between images. In this paper, we present a novel approach to lossy image compression that combines image- content classification, content-adaptive bit allocation, and hybrid wavelet tree-based coding for over 100:1 bandwidth reduction with little sacrifice in signal-to-noise ratio (SNR). Our algorithm comprises (1) content-adaptive coding that takes advantage of a classify-before-coding strategy to reduce data mismatch, (2) subimage transformation for energy compaction, and (3) a wavelet tree-based coding for efficient encoding of significant wavelet coefficients. Furthermore, instead of using the embedded zerotree coding with scalar quantization (SQ), we investigate the use of a hybrid coding strategy that combines SQ for high-magnitude outlier transform coefficients and classified vector quantization (CVQ) for compactly clustered coefficients. This approach helps us achieve reduced distortion error and robustness while achieving high compression ratio. Our analysis based on the high-frequency sonar real data that exhibit severe content variability and contain both mines and mine-like clutter indicates that we can achieve over 100:1 compression ratio without losing crucial signal attributes. In comparison, benchmarking of the same data set with the best still-picture compression algorithm called the set partitioning in hierarchical trees (SPIHT) reveals that some weak targets can completely disappear in certain situations because SPIHT is not content adaptive.

Sonar Target Recognition

The sections in this article are 1 Integrated Sonar ATR Processing 2 Real-World Experiments 3 Emerging Technologies in Sonar Target Recognition 4 Acknowledgment