Blind Signal Separation and Subspace Tracking  back

Dr. Kenneth Pope, Prof. Bob Bogner
Many signals from arrays of sensors are interdependent - e.g., microphones in a room with several conversations in progress, or sonar or radar arrays in an environment with interfering sources. The task of separating out the individual source signals is known as blind signal separation. A burgeoning interest in this area has resulted in a plethora of algorithms and applications. Work at the previous report was focused on beginning from simple, well-described signals and increasing the severity of convolutive mixing until the process breaks down. However because of the availability of many successful methods we have modified our direction of interest toward codifying the degree of difficulty of a given problem rather than simply adding additional algorithmic tricks.

Pseudoholography  back

Prof. Bob Bogner, Dr. Kenneth Pope
"Pseudoholography" was invented to answer needs in imaging of geological objects, particularly the form of ore bodies. It is potentially valuable also in imaging medical and archaeological objects, and possibly for visualisation in masonry probing. It uses virtual reality images to let humans make use contextual information and implicit constraints in the observed patterns. We process samples of data from an observed field, such as an acoustic field, to compute stereo images for the two eyes from which the forms of an object may be recognised even when the data are not sufficient to permit a complete computed reconstruction of the object. The original idea was based on holography, but from this we developed a means for reconstructing the corresponding visual images without the complications of coherent light technology.

An important issue in many real-world applications which depend on observing fields over an extended area lies in the cost of the transducer array for data gathering. Reconstruction of fields or images from sparsely-sampled fields leads to ambiguities variously referred to as aliasing or grating effects.

The equipment assembled in the previous year provides high-level control of a virtual reality headset. We used it for the initial subjective experiments and obtained apparently satisfactory 3-D images under good conditions; but this is nothing new. The previous work focusing on visualisation problems due to under-sampling of the received field was extended. The fundamental idea is to use aliased images at multiple frequencies to construct a single pseudo-hologram for visual presentation. Investigations into the use of red-green-blue colour for presentation of different frequencies has shown some perceptual benefit, but no significant breakthrough has resulted and our conclusion is that this method is not cost effective, and because only three colours are available there is only a marginal increase of dimensionality.

Various experiments probing methods of combining information at different frequencies have helped illuminate our understanding of the problem. It has emerged that the resolution of aliasing ambiguities requires use of prior knowledge. This may be through the human's implicit use of regularization in perception, or explicitly in the form of physically-imposed constraints. For example, if an object is known to be spherical, then there are few parameters and the finding of these parameters, probably via an optimization procedure is the recommended approach. Specific applications dictate what are appropriate prior assumptions for the modeling. We have the modelling of the object by an assembly of point scatterers which is identified through an optimisation algorithm.

An ARC Grant supported hardware-related parts of this work.

Signal Processing for Synthetic Aperture Radar  back

Mr. Zhishun She, Prof. Doug Gray, Prof. Bob Bogner
The project focuses on the processing of the signals in inverse SAR and multiple pass SAR. ISAR is an attractive means for providing patterns for radar target classification and recognition. It is able to reconstruct a two-dimensional high-resolution radar image of a non-cooperative target. Autofocus is a key step of ISAR imaging. Three new methods of ISAR autofocus have been developed. Their statistical performance is analysed. The Eigenvector method is identified as the maximum likelihood estimation. Cramer-Rao low bounds of ISAR autofocus and location are derived.

Multiple-pass SAR is an extension of typical interferometric SAR (InSAR). It will produce a 3D radar image of terrain. A novel algorithm for 3D image reconstruction has been developed which includes image registration, phase correction and elevational imaging. Eigenvector method, terrain centroid tracking and strong scatterer reference are developed for phase correction. 3D SAR images are obtained by processing the first European Remote Sensing Satellite (ERS-1) data with repeat orbits.

Visualisation and Pattern Recognition of Heart Rate Variability  back

Mr. Ben Raymond, Prof. Bob Bogner, Dr J. Mazumdar (University of Adelaide)
Analysis of heart rate variability (HRV) may be used as a noninvasive probe of autonomic nervous system activity. Heart rate variability analysis has been shown to have clinical application in a number of areas, including aiding the diagnosis of diabetes and better understanding the mechanisms of hypertension.

Interpretation of the information in the heart rate variability signal is not straightforward. Interpretation may be aided by the use of data visualisation techniques and, in particular, topographic mappings. Implementing the mapping in a neural network framework gains the advantage of the generalisation capability of the network, and offers the possibility of processing previously unseen data without retraining. The ability to project previously unseen data onto a mapping is an enormously useful property, allowing the comparison of unknown data to a database of known examples. In the context of medical data, this raises the possibility of use as a diagnostic aid. However, many visualisation algorithms do not offer a convenient means of projecting new data. The use of neural networks is a popular solution: once trained to produce a mapping, new data may be projected by the network with low computational cost. An investigation was made into the training of radial basis function networks for multidimensional scaling (MDS) and similar mappings. Networks may be trained by first generating a mapping of a given data set, and then training a network to reproduce it, or the two steps may be combined. It was found that, provided appropriate network regularisation is incorporated, network generalisation performance is similar for both methods. However, combining the two steps makes the construction of the mapping itself considerably easier.

The network has been demonstrated on heart rate data from a sleep apnoea study. Sleep apnoea is the cessation of breathing during sleep, and an episode of apnoea is accompanied by a characteristic heart pattern. Detection of this pattern is of clinical interest, as it may be used in automated screening for apnoea. The thesis was submitted in April 1999.

Behavior-based Reasoning Methods for Multi-Sensor Systems  back

Mr. Richard Bowyer, Prof. Bob Bogner
This research investigates methods to enhance situation awareness of an environment, that is being observed by multiple heterogeneous sensor systems. The projected outcome from this research is to enhance existing (legacy) ESM system capabilities. An interesting complication arises when the sensors are distributed and are under the control of a local authority.

The research is formulated in the context of a system of systems problem, requiring a method, or methods, for cooperation and information representation between systems. Artificial intelligence methods are being applied to the solution of this problem, with inspiration for alternative paradigms being derived from organic systems. Nature provides many examples of distributed autonomous systems, such as colonies of ants, termites and bees, that operate successfully in a cooperative manner.

This approach focuses on developing methods to integrate sensors into communities of machines that behave in a coordinated manner. The machines may be situated in close proximity to each other, or they may be widely distributed. The thesis is that machines that cooperate and have coordinated goals develop a synergy that requires little or no supervision, and will use available resources more efficiently. Further, with the application of behaviour-based reasoning (BBR) methods, an emphasis is placed on emergent behaviour that is bounded by operational constraints, and will exhibit fault tolerant behaviour.

BBR algorithms and architectures for signal and information processing are the preferred paradigm in this research. To encapsulate the ideas of a distributed processing environment, intelligent agents (IAs) are proposed as the means of information processing. The IA paradigm is used as it supports autonomy, concurrent processing, knowledge encapsulation, and adaptability. Important problems in this area of research are; action selection, mechanisms for cooperation, learning, performance evaluation, and local and global environment state representation. (ie. the local and global perceived situations.) Software Agent literature and sensor fusion literature have been reviewed this year. In the BBR approach, reactive and deliberative agents have been identified as possessing the characteristics to handle processing at different time scales.

Automatic Feature Extraction for Pattern Recognition  back

Mr. Jamie Sherrah, Prof. Bob Bogner
A typical pattern recognition system consists of two stages: the pre-processing stage to extract features from the data, and the classification stage to assign the feature vector to one of several classes. While many general classifiers exist and are well-understood, the pre-processing stage is usually ad-hoc and designed by hand.

To achieve automatic and near-optimal pre-processor design, a framework is required for the problem-independent extraction of features. Within such a framework, the concept of an optimal pre-processor can be formulated. The framework must allow pre-processors which are universally applicable and realisable using finite resources. The evolutionary pre-processor (EPrep) has been developed as a software tool to perform automatic feature extraction. EPrep performs a search over the space of generalised pre-processors, using genetic programming. EPrep has been refined and tested on 15 real and synthetic public-domain data sets. Results have shown that EPrep maintains good classification and generalisation performance. The evolved pre-processors and simple classifiers used by EPrep result in relatively accurate classification systems that can be implemented more economically than other methods.

This work has been completed.

Image Modelling and Interpretation with Application for Ionospheric Doppler Shift Correction  back

Prof. Bob Bogner, Dr David Crisp
The Jindalee OTHR is a Doppler radar and therefore requires calibration to remove the effects of Doppler shifts induced by movements of the ionosphere. This project involves the development of algorithms to do this automatically. Data is collected from the radar for selected operating frequencies and directions and is presented in the form of grey-scale images (called range-Doppler dwells) which represent the strength of the received signal as a function of range and Doppler shift. The problem is to infer the ionospheric Doppler shifts from the structures present in the dwells alone. Due to the diverse mixture of phenomena present in the dwells, modelling is not a practical solution. The techniques of pattern recognition allow an alternative approach. They allow one to design a system which is based on learning human expert interpretation of dwells from examples.

Our approach has been to develop algorithms for locating the line of zero Doppler which use simple models of the phenomena. The models are based on ideal propagation conditions and are fitted using a least squares criteria. Not surprisingly, the algorithms fail when factors not included in the models are present and consequently require supervision in their application. The fitting error has proved not to be a good measure of how well the models have explained the data. Thus we have used pattern recognition techniques to classify each range in a dwell according to how well the algorithms perform. In order to obtain training data we have implemented the algorithm in a software package which enables human experts to classify the test dwells and make any corrections necessary.

The classification is being done with Support Vector Machines (SVMs). SVMs are ideal for the problem. Their underlying principal of structural risk minimisation means they work well on sparse high dimensional data. Overcoming the curse of dimensionality was one of the main hurdles in this work. The best features for the classification task vary according to the nature of structures present in the dwell. SVMs have allowed us to work directly with the pixel data and hence avoid feature extraction problems. Our implementation of SVMs is based on John Platt's Sequential Minimal Optimisation algorithm.

The final stage of the system involves fitting the line of zero Doppler to the estimates produced by the algorithms. The fitting is done with a modified least squared error criteria. Two modifications have been used. First the fitting errors are weighted according the confidence output by the SVM classifier. Errors for estimates which are classified as poor by the SVM are given a low penalty while those for good estimates are high. Second a penalty term has been added to ensure the fitted line satisfies certain smoothness requirements.

Analysis of SAR Images back

Prof. Bob Bogner, Dr. David Crisp, Mr. Hong Gunn Chew
The task of detecting small man-made objects in Synthetic Aperture Radar (SAR) images is time consuming and repetitious. It is currently a manual process where an analyst visually processes SAR images and selects possible targets (vehicles).

Detection of man-made objects in natural surroundings may be effected by detection of significant differences between statistical aspects of the radar returns. Man-made objects may be distinguished by greater variations of brightness of reflections or by the structure of the reflections. Methods may be divided into structural and non-structural according to the degree to which spatial relationships are incorporated. Several principles are being studied relevant to different aspects.

Hypothesized targets consist of outstanding bright spots in test images. We are exploring the use of "pattern search prediction" (PSP) for locating instances of background that resemble the neighbourhoods of hypothesized targets. The essential feature of PSP is that image data from the neighbourhood of an hypothesized target are chosen as a mask and the mask is matched with neighbourhoods over a wide range of background areas. Where good matches are found the pixel values in these neighbourhoods, corresponding to the hypothesized target position in the original mask are taken as representative of the background, and are compared with the hypothesized target. The comparison is used as a basis for decisions.

If the structure of the region under consideration is not to be used then it does not matter what is the arrangement of the pixel values. Based on the idea that a man-made object should exhibit a few bright spots, it might be expected that the histogram would be clustered with a few high-level spots at the high-intensity end and with many at the low-intensity end. Such a situation corresponds to a distribution with high predictability or low entropy. In contrast, a natural background might be expected to show a more uniformly filled distribution and high entropy. We adopted the conventional information-theory approach to entropy computation and showed that the one-dimensional discriminator is not sufficient.

We are investigating the use of Support Vector Machines (SVMs) to facilitate the discrimination of targets from background and to automate parts of the target detection process. The method of SVMs represents a recently emerging aspect of learning systems with a fundamental formulation that avoids the vagaries of training in many other gradient learning approaches. SVMs have been found to perform better than traditional classifiers in many tasks. We have developed an implementation of SVMs which is based on John Platt's Sequential Minimal Optimisation algorithm. Several refinements to the original pseudo-code have been made. Our code has been written in C and is fast and accurate.

One of the advantages of SVMs is that they work well on sparse high dimensional data. This means that SVMs can be applied to the raw pixel data in the SAR images. We have produced good results for target detection with such an approach. In addition to the raw pixel data, features are also available. Other members of CSSIP's group working on this project have produced a wide range of extracted features. We have been advising them on the use of SVMs for classification. Again SVMs have performed well.

Current work has been concentrated on determining the best parameters to use in training an SVM, and on the best method for generating receiver operating characteristic curves for understanding the performance of the SVM. We are also looking at a non-linear classification technique known as Kernel Fisher Discriminant (KFD) which has grown out of the SVM theory. Fisher's Linear Discriminant has already proved to be a useful classification technique in this project. It is expected that the incorporation of kernels will be of value.

Another focus of our work in this area is the segmentation of textures. Many target detection algorithms require a uniform background to work in. In collaboration with Nick Redding at DSTO, we have been investigating segmentation algorithms based on Mumford-Shah energy functionals. Such functionals allow for a more rigorous mathematical approach to the segmentation problem. We have developed a new algorithm which we call the full lambda schedule. We have also found a useful criterion for determining the stopping condition for optimal segmentations.

The input to our texture segmentation algorithm is a stack of texture features. Good texture features are a necessary starting point. Feature extraction techniques being studied are wavelets, Gabor filters and steerable filters.