------------------------------------------------------- Statistical Signal Processing Feature-Based Classifiers ------------------------- Faculty: Stan Ahalt Students: Bill Pierson, Batu Ulug The focus of this work is the development of efficient, implementable feature-based classifiers grounded in first principles of multi-class statistical hypothesis testing. This work builds on experience and Khoros code originally developed under ARPA sponsorship. One technique which is used extensively in this work is RBF classifiers. Radial Basis Function (RBF) classifiers are efficient, parallel implementation of a Bayesian M-ary hypothesis test for feature-based target detection or identification. The list of M features forms a vector in an M-dimensional feature space; any classification scheme in effect partitions this space. The RBF approach models each target class distribution as a mixture of Gaussians. The mixture allows accurate modeling of multi-modal and non-elliptic feature parameter distributions. Restriction of the Gaussian covariance matrices serves to regularize the problem, thereby providing numerical stability and effective generalization beyond the training data. We have RBFs implemented and operating in Khoros. This work has been sponsored by WPAFB and ARPA. Hidden Markov Models -------------------- Faculty: Stan Ahalt and Ashok Krishnamurthy By exploiting anisotropic scattering, multi-aperture SAR images provide target detection that is potentially more accurate, with fewer false alarms, than normal SAR detection. Researchers have observed that man-made targets are highly anisotropic scatterers, and we exploit this anisotropic behavior in our multi-aperture HMM SAR detector. To make use of the anisotropic scattering patterns for target detection, we model ground clutter pixels, tree clutter pixels, and target pixels using discrete Hidden Markov Models (HMMs). This process uses a novel wrap-around HMM structure to faithfully model target pixels at all possible target orientations. The HMMS are trained using Baum-Welch reestimation on sample sequences representing ground clutter pixels, tree clutter pixels, or target pixels. The technique can be applied directly to traditional imagery; sub-aperture images are available without additional computational cost as part of the traditional image formation algorithms. Using what is effectively Viterbi decoding, the HMM follows the trajectory of the pixel values across the several images to statistically detect the glint phenomenon. Our preliminary results using XPatch data show that the HMM detector can be computed with fewer flops per pixel than the ubiquitous, and less effective, two-parameter CFAR detector; and performs as well or better than the TDLMS algorithm, but with greatly reduced computation. The work was sponsored by WPAFB. Data Compression: Feature set evaluation and Signature Compression --------------------------- Faculty: Stan Ahalt Students: Bill Pierson, Batu Ulug, and Jose Luis Sancho (Universidad de Madrid) We are exploring the use of compression techniques in the context of Ultra High Range Resolution radar signal returns. The goal is to determine the performance characteristics of classification systems when they are forced, because of storage limitations, to use compressed model databases. Our results show that target signatures can be reduced by factors as high as 200, with little degradation in classification performance. Using Boundary Methods, a new method for Feature Set Evaluation, we are now evaluating more extensive feature sets to determine which features support the design of robust classifiers using limited feature templates constructed from synthetic data. This work is sponsored by WPAFB and by the Air Force Maui Optical Station. Orientation Vector estimation of Satellite sub-components --------------------------------------------------------- Faculty: Stan Ahalt Students: Xun Du Extracting satellite solar-array and main-body orientation vector information from optical imagery is an integral part of Space Object Identification analysis. We are constructing a model-based image analysis system which automatically estimates the 3-D orientation vector of satellites by analyzing images obtained from ground-based optical telescopes. We adopt a two-step approach. First, pose estimates are derived from comparisons with a model database, and second, pose refinements are derived from photogrammetric information. The model database is formed by representing each available training image by a set of derived geometric primitives. To obtain fast access to the model database and to increase the probability of early successful matching, a novel indexing method is used. This work is supported by the Air Force Maui Optical Station and utilizes the resources of the Maui High Perfomance Computing Center. Image Compression ------------------ Faculty: Stan Ahalt and Raj Jain (OSU CIS) This research is directed toward image coding using Vector Quantizers developed with Artificial Neural Network (ANN) training algorithms. This work concentrates on compression techniques which are edge preserving and error-insensitive. We have constructed a unique real-time Differential Vector Quantization (DVQ) compression system which incorporates many of the techniques. This system, constructed with ASIC chips developed under sponsorship from NASA, executes in real-time, and directly addresses issues of scalable design. In related work we are developing adaptive VQ techniques and determining performance bounds on these techniques. Boundary Methods ------------------ Faculty: Stan Ahalt We are working on a collection of tools called Boundary Methods which can be used for distribution analysis, feature evaluation, informed classifier design, and Feature Set Evaluation (FSE). FSE is used to establish the rank ordering of, e.g., ATR feature sets. Using Boundary Methods, we can establish which particular feature set is most appropriate for use with a particular classifier. We are able to show, through both simulations and by analysis, that the derived ranking is consistent with alternative ranking techniques based on Bayes error. However, the use of Boundary Methods does not require extensive Monte-Carlo simulations and is not predicated on the assumption that the data is normally distributed. This work is supported by WPAFB, AFOSR, and Sverdrup Technologies, Inc. Programming Environments and Technologies ----------------------------------------- Faculty: Stan Ahalt We are supporting the DoD High Performance Computing Modernization Office (HPCMO) in an extensive upgrade of DoD computing capabilities. This project, which is focused on Signal and Image Processing (SIP) applications, has three primary objectives: 1) Deliver courses to DoD personnel which allows scientists to use HPC computers to support modern defense requirements, including: embedded HPC processors, advanced signal and image processing algorithms, image analysis and compression, and emergent parallel programming techniques. 2) Develop advanced techniques which utilize HPCs to derive new Automatic Target Recognition techniques which will be more robust to sensor noise and vehicle configuration. 3) Delivery of innovative collaborative technologies, referred to as the "Virtual Distributed Laboratory", which will allow DoD scientists to collaboratively design, prototype, test, and deploy new SIP technologies from geographically disparate laboratories, and which utilize HPC computers that are remotely located. This work is sponsored by the Army Research Lab and E-Systems, Inc.