Machine Learning (ML) is the computational study of algorithms that improve performance based on experience learned from examples. Since machine learning technique provides new learning methodologies capable of dealing with the complexities of input signals (imagery), pattern recognition investigates the applicability of modern machine learning methods to develop recognition systems with learning capabilities. This paper introduces two machine learning techniques-Algorithm Quasi-optimal (AQ) and Decision Tree (DT) as the classifiers for undertaking pattern recognition task. Both learn the 2D signal introduced from MSTAR SAR (Synthetic Aperture Radar) imagery database consisting of three classes of combat vehicles-BMP- 2, BTR-70, and T-72 tank. 67 images drawn from the database with similar aspect (+/- 15 degrees) are used for training the classifiers while unseen 47 images are used for testing. Principle Component Analysis (PCA) method and whitening transformation are used to reduce the dimensionality of input vector from 465 extracted features down to 30 features. We report three experimental results-(1) DT to learn from the original 465 features without using PCA, (2) DT algorithm with the use of PCA for reducing input dimensionality, and (3) AQ algorithm to learn the input features with PCA. The results show that the AQ has better performance than DT in terms of faster learning and higher recognition accuracy when the PCA and whitening transformation are applied.

Most Automatic Target Recognition algorithms consist of multiple processing stages, starting with a `detector' to locate objects of potential interest within an image. Then a target `classifier' identifies these objects by assigning them to specific target classes. The classifier uses the localized information in the image to assign each object to one of a number of categories, called targets, or if the object is not classifiable, it might be rejected as not being a target. This paper focuses on the properties associated with certain types of classifiers when applied to synthetic aperture radar (SAR) imagery. A common approach to classification is to construct some type of library of known templates for the targets of interest. The objects flagged by the detector are compared to each template and, based on some figure of merit, the object is classified. A popular classification rule is to calculate the mean squared error (MSE) between the detected object and each template, and assign the object to the target type that minimizes the observed MSE. Although minimization of MSE has some intuitive appeal and is fairly easy to implement, it has undesirable properties when applied to SAR data. In this paper, we investigate the statistical properties associated with MSE classification when the underlying pixel values are drawn from a long-tailed, asymmetric distribution, as is typical for SAR data. More important, however, are the within class sources of variance that arise in realistic scenarios. These sources of variance tend to inflate the MSE, even when the candidate object is compared to the correct template. This paper explores the statistical nature of this problem and illustrates it with a series of example images.

This paper introduces the ideas of zoomable laser interfaces and lenses and explains how these ideas can benefit the design of future intelligence analysis workstations.

Modem radar performs target recognition and target imaging tasks, in addition to conventional tasks of detection and tracking. New processing techniques, like stepped frequency wave-forms and RE hardware are now becoming available and will soon result in lower-cost high resolution radar for commercial as well as military applications. Advantage ofwide band operation allows generation of synthetic range with resolution of few centimeters required for target identification. An important class ofwave-forms used for high resolution mapping and target imaging, falls under the category called stretch wave-form processing. The simplest wave-form processing uses Fourier transform (FFT or IFFT). Range profiles thus generated, show the scattering centers of the target, and are being used for one—dimensional target identification procedures. These range profiles, however are very sensitive to target registration due to zero sampling inherent in the FFT procedure. This phenomenon together with the well known aspect sensitivity of the target profiles, plays havoc in the automatic target recognition procedures. In this paper we present a completely new method of obtaining range profiles. These profiles do not sample zeros and are robust with respect to range motion or range registration. Based on the super-resolution techniques, analysis is given for the sequential transform procedures. It is shown that all the peaks of the range profiles are preserved and non of the zeros are sampled. The equivalence of the present procedure to Rayleigh's Quotient is discussed. The procedure is then applied to a large set of signatures obtained using electro-magnetic code using high fidelity facet models. The range profiles were generated with the above mentioned procedures and it was found that even though there is sensitivity with respect to the aspect of the targets, the location of scattering centers remain nearly invariant for the limited aspects of the range profiles. We have designed a high dimension Bayesian classifier for the multi-class problem with empirically obtained threshold levels. The statistical separability of different classes was checked with Bhattachariyya distance for various signal to noise ratios. The dassification produces a confusion matrix and Baye's error that are close to theoretical errors for an acceptable level of signal-to-noise ratio. Results are extremely encouraging and the procedure will be extended as applied to real data.

We described a technique for using geometric information contained in 2D images to search large databases of 3D models in the special case where the geometric information consists in finite point configurations. This technique exploits certain polynomial relations known as object/image equations between invariant coordinates assigned to 2D and 3D feature sets. The resulting scheme is invariant to changes in scale and perspective. Here, we describe a technique for constructing indexes (i.e. a hashing scheme) based on this technology.

A new technique for the detection of oil leaks in a power plant is presented which is based on artificial neural networks. In this system the neural network was trained by feature parameters extracted from ITV images of the normal condition and oil leaks. For input data in the neural network we used four parameters calculated from two regions of interest in a subsequent image. They were the rate of variation of pixel values, variance of the pixel values, skewness of the pixel values and rate of variation of the skewness. The results showed that the accuracies of recognition were more than about 90%. The system is considered to be helpful for industrial surveillance application.

The volume of video and other image data is expanding at a rapid pace with the increasing use of surveillance systems, unmanned vehicles, and other collection systems. The sheer volume of images requires the use of automatic systems to select interesting image features for further analysis. These systems should have a low false alarm rate, e.g. satisfying a pre-determined constant false alarm rate (CFAR). Various filters may be applied to filter out non- target (background) parts of an image. The output of these filters is noise, plus possible target features. When the noise is Gaussian, CFAR thresholds may be based on t- distributions, with reduced degrees of freedom in the case of correlated noise. For the non-Gaussian case, the use of t distributions is inappropriate, and we suggest alternatives based on parametric families of distributions, with location, scale, and shape parameters. When shape parameters are known the thresholds can be determined using a Monte Carlo technique, using variance reduction techniques to improve the computational efficiency by a factor of 1800. We discuss methods for handling unknown shape parameters.

In this paper, we study the problem of estimating and segmenting the optical flow field in image sequences. A variational framework based on the Mumford-Shah functional is introduced for simultaneous edge preserved optical flow estimation and motion-based segmentation. The proposed energy functional for optical flow field and its corresponding edge set if formulated to have three additive terms. The first and second terms measure the deviation from the optical flow constraints over the whole image and its smoothness at all the non-edge locations in L₂ norm, respectively, while the third term regularizes the total length of all the edges. The minimization of this functional is carried out by the vector graduated nonconvexity (VGNC) algorithm with the gradient descent iterating scheme. This framework is then extended to fuse spatio-temporal segmentation by adding two more terms for spatial segmentation in the above formulation. One term is the L₂ difference between the original image and an approximation of the image, while the other is the regularization of approximate image at all non-edge locations. The same VGNC procedure is performed to minimized the functional to obtain the optical flow field, the piecewise smooth image, and the spatio-temporal edge image. We illustrate the presented method and its numerical implementation on tactical image sequences.

Numerous Optical Character Recognition (OCR) companies claim that their products have near-perfect recognition accuracy (close to 99.9%). In practice, however, these accuracy rates are rarely achieved. Most systems break down when the input document images are highly degraded, such as scanned images of carbon-copy documents, documents printed on low-quality paper, and documents that are n-th generation photocopies. Besides, the end user cannot compare the relative performances of the products because the various accuracy results are not reported on the same dataset.. In this article we report our evaluation results for two popular Arabic OCR products: (1) Sakhr OCR and (2) OmniPage for Arabic. In our evaluation we establish that the Sakhr OCR product has 15.47% lower page error rate relative to the OmniPage page error rate. The absolute page accuracy rates for Sakhr and Omnipage are 90.33% and 86.89% respectively. Our evaluation was performed using the SAIC Arabic image dataset, and we used only those pages for which both OCR systems produced output. A scatter-plot of the page accuracy-rate pairs reveals that Sakhr in general performs better on low-accuracy (degraded) pages. The scatter-plot visualization technique allows an algorithm developer to easily detect and analyze outliers in the results.

Static signature verification is a well researched problem that has not been completely solved to this date. To improve on current verification performance this research uses a pooling method which fuses together decisions of selected verification algorithms. To enhance this performance further, the decision from this method is fused with the decision of a neural network classifier. This neural network classifier offers a new approach to signature verification, since it is based on recognition techniques. The advantage of this classifier is that it incorporates different information into its decision and therefore allows the fused decision to be based on more diverse information. In contrast to other methods, this classifier requires only genuine signature samples to be trained. Experimental results show that the fusion of verification algorithms can produce better performance than any of the used methods individually.

We present a language-independent optical character recognition system that is capable, in principle, of recognizing printed text from most of the world's languages. For each new language or script the system requires sample training data along with ground truth at the text-line level; there is no need to specify the location of either the lines or the words and characters. The system uses hidden Markov modeling technology to model each character. In addition to language independence, the technology enhances performance for degraded data, such as fax, by using unsupervised adaptation techniques. Thus far, we have demonstrated the language-independence of this approach for Arabic, English, and Chinese. Recognition results are presented in this paper, including results on faxed data.

Recent work has indicated that polarization difference imaging has the potential to enhance the image quality of objects viewed in the presence of scattering media, such as fog and turbid water. We have utilized an AOTF spectro- polarimeter to implement this concept, and to expand its usefulness by incorporating real-time, adaptive, complex polarization and spectral filtering techniques into the system.

The paper presents results on theoretical and experimental investigation of tunable acousto-optic filters and their applications in optics and spectroscopy for spectral filtration of divergent optical beams and images. Physical principles of operation of wide-angular acousto-optic filters are examined and basic parameters of developed instruments are described in the paper. The instruments intended for processing of optical images are designed on base of optically anisotropic single crystals of tellurium dioxide. Various experimental block-schemes and optical arrangements for image processing are described and some peculiarities of the imaging experiments are presented in the paper. In particular, experimental results on acousto- optic imaging by means of tunable filters not sensitive to polarization of incident optical radiation are examined and processing of infrared images characterized by relatively low intensity of light beams is described in the report. Proposals on design of acousto-optic cells optimized for the imaging experiments are also presented and discussed in the paper.

Acousto-optic tunable-filter (AOTF) technology has been used in the design of a no-moving parts, compact, lightweight, field portable, automated, adaptive spectral imaging system when combined with a high sensitivity imaging detector array. Such a system could detect spectral signatures of targets and/or background, which contain polarization information and can be digitally processed by a variety of algorithms. At the Army Research Laboratory, we have developed and used a number of AOTF imaging systems and are also carrying out the development of such imagers at longer wavelengths. We have carried out hyperspectral and multispectral imaging using AOTF systems covering the spectral range from the visible to mid-IR. One of the imager uses a two-cascaded collinear-architecture AOTF cell in the visible-to-near-IR range with a digital Si charge-coupled device camera as the detector. The images obtained with this system showed no color blurring or image shift due to the angular deviation of different colors as a result of diffraction, and the digital images are stored and processed with great ease. The spatial resolution of the filter was evaluated by means of the lines of a target chart. We have also obtained and processed images from another noncollinear visible-to-near-IR AOTF imager with a digital camera, and used hyperspectral image processing software to enhance object recognition in cluttered background. We are presently working on a mid-IR AOTF imaging system that uses a high- performance InSb focal plane array and image acquisition and processing software. We describe our hyperspectral imaging program and present results from our imaging experiments.

Hyperspectral imagery is currently being evaluated for its potential to solve a number of complex imagery intelligence issues. Acousto-Optical Tunable Filters offer a potentially valuable alternative to standard imaging spectrometer designs. The work from previous and on-going programs is indicative of this potential.

The acousto-optic tunable filter (AOTF) has a narrow passband and a large angular acceptance angle, which allows for imaging at a given wavelength without having to assemble an image cube, as with grating based imagers. It is also possible to use an acousto-optic dispersive filter (AODF), which has a small acceptance angle and a broad spectral passband to form spatial images similar to the grating imager. Although the processing is more complex, the advantages of the AODF are that pixel registration and temporal fluctuations of the spectra are greatly reduced compared to the AOTF. Both the AOTF and AODF can operate in a birefringent mode, allowing for the use of high efficiency materials such as Tl₃AsSe₃ (TAS) in the infrared region. They can both be multiplexed to increase the sensitivity, and reduce the spectral fluctuation problem of the AOTF. The AODF can also operate in an isotropic mode, which allows for the use of deflector materials such as Ge. The issues of complexity, fluctuations, efficiency, and multiplexing are compared for AOTFs and AODFs operating in the infrared. A comparison is also made for both systems using TAS, along with AODFs using Ge.

By utilizing the unique AOTF feature of varying its spectral transmission according to the spectrum of the driving RF signal, Physical Optics Corporation developed an automatic hyperspectral system for real-time target contrast enhancement and target tracking.

In this paper, we propose our image indexing technique and visual query processing technique. Our mental images are different from the actual retinal images and many things, such as personal interests, personal experiences, perceptual context, the characteristics of spatial objects, and so on, affect our spatial perception. These private differences are propagated into our mental images and so our visual queries become different from the real images that we want to find. This is a hard problem and few people have tried to work on it. In this paper, we survey the human mental imagery system, the human spatial perception, and discuss several kinds of visual queries. Also, we propose our own approach to visual query interpretation and processing.

This paper presents the result of our research on a new associative memory, which unlike any existing neural network based artificial associative memories, can dynamically localize (or focus) its search on any subset of the pattern space. This new ability now makes the power of associative computing available to a new class of pattern matching applications. Application areas which will particularly benefit from this model include (1) detection of small irregular patterns (medical diagnostics), (2) detection of tiny targets, (3) background varying target recognition, (4) visual example based content-based image retrieval, (5) robust adaptive control systems which needs to continue operating with small number of surviving sensors, in the face of post learning loss of sensors.

ALISA (Adaptive Learning Image and Signal Analysis) is an adaptive statistical learning engine that may be used to detect and classify the surfaces and boundaries of objects in images. The engine has been designed, implemented, and tested at both the George Washington University and the Research Institute for Applied Knowledge Processing in Ulm, Germany over the last nine years with major funding from Robert Bosch GmbH and Lockheed-Martin Corporation. The design of ALISA was inspired by the multi-path cortical- column architecture and adaptive functions of the mammalian visual cortex.

We are currently exploring the relationship between spatial statistical parameters of various geophysical phenomena and those of the remotely sensed image by way of principle component analysis (PCA) of radar and optical images. Issues being explored are the effects of incorporating PCA into land cover classification in an attempt to improve its accuracy. Preliminary results of using PCA in comparison with unsupervised land cover classification are presented.

The need for robust image data sets for algorithm development and testing has prompted the consideration of synthetic imagery as a supplement to real imagery. The unique ability of synthetic image generation (SIG) tools to supply per-pixel truth allows algorithm writers to test difficult scenarios that would require expensive collection and instrumentation efforts. In addition, SIG data products can supply the user with `actual' truth measurements of the entire image area that are not subject to measurement error thereby allowing the user to more accurately evaluate the performance of their algorithm. Advanced algorithms place a high demand on synthetic imagery to reproduce both the spectro-radiometric and spatial character observed in real imagery. This paper describes a synthetic image generation model that strives to include the radiometric processes that affect spectral image formation and capture. In particular, it addresses recent advances in SIG modeling that attempt to capture the spatial/spectral correlation inherent in real images. The model is capable of simultaneously generating imagery from a wide range of sensors allowing it to generate daylight, low-light-level and thermal image inputs for broadband, multi- and hyper-spectral exploitation algorithms.

The Support Vector Machine provides a new way to design classification algorithms which learn from examples (supervised learning) and generalize when applied to new data. We demonstrate its success on a difficult classification problem from hyperspectral remote sensing, where we obtain performances of 96%, and 87% correct for a 4 class problem, and a 16 class problem respectively. These results are somewhat better than other recent result on the same data. A key feature of this classifier is its ability to use high-dimensional data without the usual recourse to a feature selection step to reduce the dimensionality of the data. For this application, this is important, as hyperspectral data consists of several hundred contiguous spectral channels for each exemplar. We provide an introduction to this new approach, and demonstrate its application to classification of an agriculture scene.

Digital elevation models provide an estimate of the topographic fabric (the tendency for topography to have a preferred orientation) of the earth's surface. The algorithm extracts the eigenvectors and eigenvalues from a 3 X 3 matrix of the sums of the cross products of the directional cosines of the surface normals computed at each point in the DEM. The ratio of eigenvalues S₁ (largest) and S₂ measures the ruggedness of the terrain. The ratio of eigenvalues S₂ and S₃ (smallest) measures the tendency for the terrain to have a preferred orientation, while their orientation reflects the direction of dominant topographic fabric. Sample sizes of about 500 - 2500 points provide robust statistics, allowing sample regions of 1/2 to 1/3 square degree for global data sets and about 600 meters on a side with 30 m US topography. Topographic fabric appears to be a fundamental characteristic of landforms amenable to quantitative study. It should be included in terrain analysis and classification, and may lead to better estimates for cross-country mobility.

As part of collaboration between the Pittsburgh Supercomputing Center and the University of Pittsburgh Medical Center we are developing methods for content based image retrieval to assist pathology diagnosis. We have been using Gleason grading of prostate tumor samples as an initial domain for evaluating the effectiveness of the method for specific tasks. In this application, the system does not attempt to directly reproduce pathologists' visual analysis. Rather, it relies on the comparison of image features from a sample image to key the retrieval of similar but previously graded images from a database. Appropriate features should be highly selective to architecture differences of the Gleason system so the grades of the retrieved images can be applied to the unknown sample. We have been investigating the usefulness of computational geometry structures, such as spanning trees, as components of feature sets providing accurate retrieval of matching grades.

Object description is important for performing photon transport efficiently by means of a Monte Carlo method. The description methods include a voxel-based description which represents an object by means of the union of a voxel of the same size and an octree description which describes an object by using cubic regions of several sizes. The octree representation requires fewer regions than the voxel-based description when the object is represented with the same precision. Therefore the octree representation very effectively reduces the calculation time, but the number of regions depends on the coordinate system of the octree representation. We investigated the relationship between the description methods and the performance of the calculation by using each description. We calculated the photon transportation by using the object structure determined in different coordinate systems. The results showed that the calculation speed depends upon the coordinate systems used for representing the phantom even though the number of regions does not change.

We have previously presented a hierarchical pyramid/neural network (HPNN) architecture which combines multi-scale image processing techniques with neural networks. This coarse-to- fine HPNN was designed to learn large-scale context information for detecting small objects. We have developed a similar architecture to detect mammographic masses (malignant tumors). Since masses are large, extended objects, the coarse-to-fine HPNN architecture is not suitable for the problem. Instead we constructed a fine-to- coarse HPNN architecture which is designed to learn small- scale detail structure associated with the extended objects. Our initial result applying the fine-to-coarse HPNN to mass detection are encouraging, with detection performance improvements of about 30%. We conclude that the ability of the HPNN architecture to integrate information across scales, from fine to coarse in the case of masses, makes it well suited for detecting objects which may have detail structure occurring at scales other than the natural scale of the object.