Our main objective is the integration of sensor physics and the tools of statistical analysis and modeling into Image Understanding (IU) methods and techniques. We believe that a precise analysis of sensor physics, performance and noise will lead to new models and approaches that suggest novel ways in which IU concepts need to be adapted and extended in order to make optimal use of all available data. The set of interrelated research topics that will serve as the initial focus for our research include:

• Multiresolution SAR-based ATR
• SAR-based ATR incorporating pose-dependent SAR image formation and analysis

Our approach involves a careful integration of sensor physics, statistical analysis and modeling, and IU methods. In particular, we plan to take the following approach to each of the interrelated topics.

One of our major successes to date under the current ATR-URI project is in developing multiresolution stochastic models for SAR imagery that accurately capture the scale-to-scale statistical variability of speckle in SAR imagery. In one application, we used models for natural clutter and for man-made objects together with our fast statistical likelihood calculation methods to develop an enhanced discrimination feature that, when integrated into Lincoln Labs' ATR algorithm and tested on a very large data set, resulted in a factor of 6 reduction in false alarms over the previous best results. In the second application, we used our models to segment natural clutter (trees and grass) and to enhance anomalous pixels (due to man-made scatterers) that did not produce the scale-to-scale variability consistent with natural clutter. The results are very accurate segmentations and enhanced visibility of anomalies as compared to widely used CFAR methods.

We believe there are many additional applications for these multiscale models. For example, anomalies that result from man-made objects exhibit themselves as distinctive patterns across scale that differ significantly from the scale-to-scale textural variations. Consequently, chains of pixels across scale could in principle be viewed as robust and statistically meaningful features that can be further exploited for model-based recognition. We propose to use multiscale features for higher level recognition and reasoning. Classically, target models include geometrical constraints on the appearance of features in space. In this new framework, models will also include information about the appearance of features across scale. The development of such models is a central objective of this project. Once we have such models, we can use our statistically optimal methods for evaluating likelihoods to evaluate match scores for hypothesized models and poses.

A widely recognized fact in the ATR community is that the scattering patterns for man-made scatterers possess very different characteristics from those of natural scatterers. In particular, while the latter frequently can be modeled as diffuse isotropic scatterers, the former frequently have strong specular characteristics, which implies that they have extremely strong aspect-dependent responses. This effect is even more pronounced in low-frequency SAR as is used for foliage penetration.

While some work has attempted to account for the difference in scatterer type, it is fair to say that a fundamental look at this problem has yet to be taken. In some very recent research Prof. Shapiro has begun to do this. His analysis indicates that there is significant discriminating information to be obtained if one examines sets of SAR images of a scene constructed using different subapertures of the full SAR aperture. What this suggests is that the natural data structure for higher-level ATR functions will again involve imagery at different resolutions, but that in this case the focus is on exploiting differences in imagery obtained from different parts of the aperture (and hence from different viewing angles). Note that this results in a very novel pose estimation problem: the model of a specular reflector must capture the fact that changes in pose not only change the relative geometry of features but also can change the appearance of these features in imagery from different apertures. As a result, the action of the pose transformation group on a set of images from different apertures and at different resolutions is not simply a geometric transformation.

We are still working out the details of the demonstrations, which will include:

• Develop multiscale statistical models of SAR imagery with applications in multiscale texture modeling and the definition of a methodology for constructing multiscale models for SAR images of targets. This includes defining pose transformation operations on such multiscale representations
• Develop detailed statistical models for SAR data for both man-made and natural scatterers. The objective here is to pinpoint statistically significant difference due to aspect dependencies
• Define a methodology to construct and model multiresolution imagery from several subapertures in order to capture the aspect-dependent differences identified in our research.

Abstracts of current research projects supported under this grant.

Project Abstracts	Structure Driven SAR Image Registration Jeremy S. De Bonet Flexible Histograms: A Multiresolution Texture Discrimination Model Jeremy S. De Bonet Multiresolution Sampling Procedure For the Analysis And Synthesis Of Texture Images Jeremy S. De Bonet Noise Reduction Through Detection of Signal Redundancy Jeremy S. De Bonet

Other projects at the MIT AI lab which are related to the goals of this grant.

Aug 1997

Dec 1997

Apr 1998

Jul 1998

Oct 1998

Jan 1999

Apr 1999

Jul 1999

Oct 1999

Jan 2000

IMEX Review June 1998, Wright Patterson AFB

MSTAR Briefing July 1998 (1)

MSTAR Briefing July 1998 (2)

Image Understanding Workshop 1998, CA