Quyen Q. Huynh

Classification of underwater mammals using feature extraction based on time-frequency analysis and BCM theory

Underwater mammal sound classification is demonstrated using a novel application of wavelet time-frequency decomposition and feature extraction using a Bienenstock, Cooper, and Munro (1982) (BCM) unsupervised network. Different feature extraction methods and different wavelet representations are studied. The system achieves outstanding classification performance even when tested with mammal sounds recorded at very different locations (from those used for training). The improved results suggest that nonlinear feature extraction from wavelet representations outperforms different linear choices of basis functions.

Sonar object discrimination via spectral density analysis

Current underwater sonar exploration often neglects the information that exists in the spectral density of object returns. The attempt to perform spectral density exploration of objects is complicated by the difficulty to present informative spectral content of each location pixel on a 2D image. Spectral content is best represented by the spectral density; however, it does not make sense to explore the detailed spectral density for each pixel in the image. In this work, we outline a general approach for spectral density analysis which is intended for enhancing object discrimination and delivers an easy to interpret, image enhancement of sonar returns. This method actually enhances the sonar image with acoustic color which emphasizes an optimal combination of frequency bands of the returned spectrum for the purpose of object discrimination. This is achieved by analyzing the discriminating power of different frequency bands and creating an optimal association with different bands to a corresponding color map.

Image enhancement for pattern recognition

We investigate various image enhancement techniques geared towards a specific detector. Our database consists of side- scan sonar images collected at the Naval Surface Warfare Center (NSWC), and the detector we use has proven to have excellent results on these data. We start by investigating various wavelet and wavelet packet denoising methods. Other methods we consider are based on more common filters (Gaussian and DOG filters). In wavelet based denoising we try different approaches, combining techniques that have been successfully used in signal and image denoising. We notice that the performance is mostly affected by the choice of the scale levels to which shrinkage is applied. We demonstrate that wavelet denoising can significantly improve detection performance while keeping low false alarm rates.

Feature extraction and fusion of wideband backscattered signals

Good discrimination result have been obtained with an active backscatter data set of mine-like objects, where the task as to distinguish between man-made and non-man-made objects. In this work we introduce a novel method for constructing best basis for discrimination from wavelet packets, and demonstrate the superiority of multiple ensembles of predictors. We achieve far better discrimination results using a wide band FM sweep of 80kHz compared with the earlier work that used a 40kHz FM sweep. We further show improved results by combining several discriminative methods with several wavelet packet representations.

GINOM: A statistical framework for assessing interval overlap of multiple genomic features

A common problem in genomics is to test for associations between two or more genomic features, typically represented as intervals interspersed across the genome. Existing methodologies can test for significant pairwise associations between two genomic intervals; however, they cannot test for associations involving multiple sets of intervals. This limits our ability to uncover more complex, yet biologically important associations between multiple sets of genomic features. We introduce GINOM (Genomic INterval Overlap Model), a new method that enables testing of significant associations between multiple genomic features. We demonstrate GINOM’s ability to identify higher-order associations with both simulated and real data. In particular, we used GINOM to explore L1 retrotransposable element insertion bias in lung cancer and found a significant pairwise association between L1 insertions and heterochromatic marks. Unlike other methods, GINOM also detected an association between L1 insertions and gene bodies marked by a facultative heterochromatic mark, which could explain the observed bias for L1 insertions towards cancer-associated genes.

Seafloor identification in sonar imagery via simulations of Helmholtz equations and discrete optimization

Abstract We present a multiscale approach for identifying features in ocean beds by solving inverse problems in high frequency seafloor acoustics. The setting is based on Sound Navigation And Ranging (SONAR) imaging used in scientific, commercial, and military applications. The forward model incorporates multiscale simulations, by coupling Helmholtz equations and geometrical optics for a wide range of spatial scales in the seafloor geometry. This allows for detailed recovery of seafloor parameters including material type. Simulated backscattered data is generated using numerical microlocal analysis techniques. In order to lower the computational cost of the large-scale simulations in the inversion process, we take advantage of a pre-computed library of representative acoustic responses from various seafloor parameterizations.

Compression for small targets in multispectral imagery

This paper analyzes the performance of a fast, low complexity, integer-to-integer compression scheme that is designed to give greater importance to small targets. Practical real-time operation of unmanned aerial vehicle mine/minefield detection systems has two difficult constraints. One is limited data-link bandwidth and the other is limited on-board processing power. Standard compression techniques are usually complex and tend to remove small objects from the imagery. In the imagery used for airborne mine/minefield detection, the targets are small, usually on the order of a few pixels. The region-of-interest (ROI) Wavelet Difference Reduction (WDR) compression scheme satisfies both of these con-straints and is shown to preserve detection rates of small targets. Results are compared for block-based (BB)-WDR compressed and ROI-WDR compressed and uncompressed images. The ROI -WDR process is shown to be superior to other compression conditions.

Sonar image enhancement via acoustic color

A novel method to enhance synthetic aperture sonar (SAS) images using the acoustic color of the returns from different targets is presented. The method enhances the SAS image using the internal absorption properties of targets in a way that enhances the differences between different types of targets (target signature). First, the algorithm detects the highlights in the sonar image as is found from the beam‐formed processed image. These highlights represent edges of targets. Then the analysis is performed only at the echo returns which correspond to edges of targets. For each edge the raw data corresponding to the return from that edge is found. This data which has not yet been pulse‐compressed using the pinging signal is analyzed using a detailed frequency representation. Then, the frequency representation of the pinging signal is subtracted from it so that a resulting frequency representation of the difference between the pinging signal and the returning signal is obtained. A normalization of the returne...

Target reconstruction from broadband acoustic backscatter

Broadband acoustic backscatter data are examined for mine reconstruction using LFM and dolphin clicks. Broadband backscatter data from free‐field and bottom mines were collected during tests conducted in ARL/UT’s sand tank and sponsored by ONR. The data set includes multi‐aspect broadband LFM (50–200 kHz) acoustic backscatter returns from a collection of targets. Both free‐field and bottom target data were collected. Bottom target data sets include returns from two different levels of bottom roughness. Each data file is sampled over an interval that is large relative to the echo duration with sufficient sampling both prior to the echo onset and post‐echo return to enable meaningful target characterization. Data collected over 360 deg in 2‐deg increments are processed to generate a descriptive target pattern. The data are pulse compressed and system artifacts are removed using adjacent channels over the entire frequency band and within subbands. The target pattern is generated by applying the new adaptive ...

Time‐frequency analysis of model based pulse scattering from a rough fluid‐elastic interface

A null field T‐matrix formalism for plane‐wave scattering from a doubly infinite fluid–solid interface with doubly periodic surface roughness [Bishop and Smith, J. Acoust. Soc. Am. 94, 1560–1583 (1993)] is extended to include scattering of broadband acoustic pulses. The scattered pressure field is expressed in terms of a Fourier integral over the spectra of the incident pulse and the T matrix. Time series are calculated for the scatter of a Gaussian amplitude‐modulated linear frequency‐modulated pulse from a fluid–elastic interface with sinusoidal and triangular roughness. The time series are analyzed using short time Fourier, Wigner, wavelet packet, and local cosine T‐F transforms. The results are displayed graphically in a T‐F plot in which the T‐F plane is covered by rectangular boxes of dimensions Δt×Δω. These results are used to evaluate the ability of each of the transforms to address inverse scattering and identification problems. It is shown that the T‐F analysis provided by the wavelet packet may...