In this paper we present a modification of the test phase of ARTMAP-based neural networks that improves the classification performance of the networks when the patterns that are used for classification are extracted from noisy signals. The signals that are considered in this work are textured images, which are a case of 2D signals. Two neural networks from the ARTMAP family are examined, namely the Fuzzy ARTMAP (FAM) neural network and the Hypersphere ARTMAP (HAM) neural network. We compare the original FAM and HAM architectures with the modified ones, which we name FAM-m and HAM-m respectively. We also compare the classification performance of the modified networks, and of the original networks when they are trained with patterns extracted from noisy textures. Finally, we illustrate how combination of features can improve the classification performance for both the noiseless and noisy textures.

We present here a mammography imaging system, named DAF Sparkle^TM, with a particular emphasis on digital mammogram enhancement and feature analysis using interpolating Distributed Approximating Functional (DAF) wavelets, nonlinear multiscale edge enhancement, and visual group normalization technique (VGN). Sinc-type interpolating DAFs are utilized for generating a new class of biorthogonal wavelets with arbitrary smoothness in both space and frequency regions. A new nonlinear multiscale edge enhancement technique is presented here to sharpen the image edge based on the wavelet coefficients without enhancing the noise as well. The visual group normalization technique (VGN), as a natural extensions of earlier normalization techniques for noisy image restoration, is used to normalize the multiscale wavelet coefficients, remove perceptual redundancy, as well as to improve the visualization of the important diagnostic features for digital mammograms. Without prior knowledge of the `true' spatial distribution of the signal, excellent enhancement results are obtained by the present techniques.

In this paper, a texture based algorithm is developed for classifying color images. The images are filtered by a set of Gabor filters at different scales and orientations. The energy of the filtered images in each channel and between channels are computed and used for classification. The normalized RGB, xyY and HIQ color spaces are used to identify the best space for classifying the color images. The best representation of the textures are found to be using normalized RGB and HIQ space and chrominance values. A filter selection process using texture similarity is adopted. Unichannel and interchannel features are computed. A feature reduction process is applied before using a classifier. The algorithm is used to classify sets of textures from databases of color texture images and it gives good results. It is also applied to Landsat TM images. The 7 channels are used and the best channels for classification of the image are found to be R and G. The algorithm has been designed to use the appropriate Gabor filters based on texture transition characteristics within and between channels. The algorithm performs better than using only the gray scale values of the color images.

Image preprocessing is useful in helping to identify `spectral response patterns' for certain types of image classification problems. The common artifacts in remotely sensed images are caused by the blurring due to the optics of the image gathering device, illumination variations, and the radiative transfer of the atmosphere. The Multi-Scale Retinex (MSR) image enhancement algorithm that provides dynamic range compression, reduced dependence on lighting conditions, and improved (perceived) spatial resolution has proven to be an effective tool in the correction of image degradations such as those in remote sensing images. In this paper, we measure the improvement in classification accuracy due to the application of the MSR algorithm. We use simulated images generated with different scene irradiance and with known ground truth data. The simulation results show that, despite the degree of image degradation due to changes in atmospheric irradiance, classification error can be substantially reduced by preprocessing the image data with the MSR. Furthermore we show that, similar to the results achieved in previous work, the classification results obtained from the MSR preprocessed images for various scene irradiance are more similar to each other than are the classification results for the original unprocessed images. This is evident in the observed visual quality of the MSR enhanced images even before classification is performed, and in the different images obtained by comparing image data under different irradiance conditions. We conclude that the application of the MSR algorithm results in improved visual quality and increased spatial variation of multispectral images that is also optimal for certain types of multispectral image classification.

We applied dithered quantization to image compression using a wavelet transform, scalar quantization method. The results indicate that dithered quantization could change the noise characteristics of the reconstructed image.

This paper presents a low-complexity method of transmitting digitally compressed imagery through AWGN and fading channels. The proposed method combines a wavelet-based image coder that employs phase scrambling and trellis-coded quantization (TCQ), and source adaptive modulation (SAM). We present two versions of SAM that utilize BPSK (SAM-TCQ) and 16-ary PPM (OSAM-TCQ), respectively. We then compare the performance of the SAM systems to that of a system that utilizes unequal error protection (UEP). We show that for the binary symmetric channel, the SAM-TCQ system performs as well as UEP-TCQ at high bit error rates, and within 1 dB of the UEP-TCQ system at low bit error rates. Additionally, we show that for the AWGN channel, the OSAM-TCQ system performs nearly 4 dB better than UEP-TCQ at high bit error rates, and the same as the UEP-TCQ system at low bit error rates, with much lower complexity than the UEP-based system.

In lossy image compression schemes, often some distortion measure is minimized to arrive at a desired target bit rate. The distortion measure that has been most studied is the mean-squared-error (MSE). However, perceptual quality often does not agree with the notion of minimization of mean square error1 . Since MSE can not guarantee the optimality of perceptual quality, others error measures have been investigated. Others have found strong mathematical and practical perspective to choose a different error measure other than MSE, especially for image compression2. In Ref. 2 it is argued that the mean absolute error (MAE) measure is a better error measure than MSE for image compression from a perceptual standpoint. In addition, the MSE measure fails when only a small proportion of extreme observations is present3. In this paper we develop a bit allocation algorithm to minimize the MAE rather than MSE

Coordinate calibration is a critical step in reconstructing 3D range images from 2D camera images of structured light projections. Having accurate reconstruction yields accurate measurement of objects, assembly of multiple views of the same object from different directions and interlacing of multiple camera and projection field of views. We present a robust calibration algorithm, which combines a 3D registration algorithm with structure light projections. The light projection is implemented with a spatial light modulator using an array of deformable mirrors. The approach uses 6 or more non-coplanar registration points on a physical calibration device. These points are in view to both the camera and projector. The camera and projector orientations are unknown. Singular value decomposition is applied to this data to estimate the perspective transformation parameters between 3D world coordinate and 2D camera coordinates. The same process is applied to find the transformation parameters between the projector coordinates and the world coordinates. From these two transformations we incorporate the lateral distortions produced by structured light projections and reconstruct manifolds that accurately represent the surfaces being measured. While sine wave projection is used in this study, the calibration and reconstruction methodology can be applied to a variety of structured light projection methods. Experimental results demonstrate the algorithm performance.

The U.S. Army Research Laboratory (ARL) is investigating an efficient SAR image formation algorithm to support its mission-funded Ultra-Wideband BoomSAR program. The traditional Back-projection technique produces images, requires no memory resources and allows arbitrary motion; however it is very computationally intensive and is therefore appropriate only for postprocessing applications. ARL invented a new recursive backprojector that reduces the computational load from an order N³ to N²Log(N). This new algorithm is competitive in speed with the frequency domain WK processing but requires less memory, and produces fewer artifacts. We compare both algorithms and presents the results of using simulation data and real data.

Charge coupled-device imaging systems are often designed so that the image of the object field is sampled well below the Nyquist limit. Undersampled designs frequently occur because of the need for optical apertures that are large enough to satisfy the detector sensitivity requirements. Consequently, the cutoff frequency of the aperture is well beyond the sampling limits of the detector array, and aliasing artifacts degrade the resulting image. A common antialiasing technique in such imaging systems is to use birefringent plates as a blur filter. The blur filter produces a point spread function (PSF) that resembles multiple replicas of the optical system PSF, with the separation between replicas determined by the thickness of the plates. When the altered PSF is convolved with the PSF of the detector, an effective pixel is produced that is larger than the physical pixel and thus, the higher spatial frequency components and the associated aliasing are suppressed. Previously, we have shown how information theory can be used in designing birefringent blur filters by maximizing the information density of the image. In this paper, we investigate the effects of spherical aberration and defocus on the information density of an imaging system containing a birefringent blur filter.

The size and dimensions of a focal plane array do not necessarily limit the achievable resolution of a digital imaging camera if such a camera can make repeated exposures of a scene where all such exposures differ by some perspective transformation. This paper outlines a generalized reconstruction approach that does not depend on controlled micro dithering of the camera, nor requires the set of exposures to maintain a strictly uniform translation.

A commercial version of the 3D Artificial Neural Network has been developed under a collaborative effort between JPL and Irvine Sensors, sponsored by BMDO and the U.S. Air Force. It is capable of continuous trillion eight bit multiply and add operations per second while consuming under ten watts. Its architecture, input-output characteristics, and performance data will be presented.

Many of the limitations of traditional optical-only imaging systems can be eliminated with jointly optimized optical and digital imaging systems. Jointly optimized optical and digital imaging systems exploit the complementary aspects of optics and digital signal processing to form systems with characteristics not possible with traditional optics-only systems. For example, in traditional imaging systems light gathering and large depth of field are competing goals and are inversely related. On the other hand, in optimized optical/digital imaging systems light gathering and large depth of field can be independent parameters. Instead of requiring a small aperture to produce a large depth of field, a large aperture and a large depth of field are both possible and practical. We can jointly optimized optical and digital imaging systems Wavefront Coded imaging systems. Concepts of Wavefront Coding are illustrated below through an athermalized, refractive, silicon/germanium IR imaging system with aluminum optical mounts subject to an ambient temperature range of -20 degree(s)C to +70 degree(s)C.

The future of remote sensing includes a significant role for hyperspectral imaging. Hyperspectral imaging can provide the data needed to derive detailed information on composition, biomass health, military status and other properties of Earth's surface and atmosphere. The biggest challenges currently associated with hyperspectral imaging are related to handling high data rates in timely and efficient ways. This paper introduces hyperspectral imaging, discusses systems that have been built and tested and describes future system concepts.

The effectiveness of utilizing spatial light modulators (SLMs), developed at Sanders, for reducing some of the critical bottlenecks inherent within the Hyperspectral Imaging (HSI) chain will be presented. Specifically, the more common classification, detection, and endmember selection algorithm used in HSI, which are suitable for optical implementation, are presented here. These algorithms were reformulated for implementation on a compact Vander- Lugt correlator based on Sanders' multi-level quantum well (MQW) spatial light modulator (SLM). Sanders devices are GaAs Fabry-Perot vertical cavity multiple quantum well (MQW) SLMs consisting of MQW optical chips flip-chip bonded to Si/CMOS driver circuitry. Details of the reformation of Pixel Purity Index, an endmember selection algorithm, to the optical correlator is presented as well as a projection of its performance based on software simulations. In addition, hardware results are presented for Spectral Angle Mapper based on a Vander-Lugt implementation using Sanders 128 X 128 binary SLMs. An opto-electronic hyperspectral workstation accelerator is proposed which is based on a Vander-Lugt correlator using Sanders' MQW-SLMs and FPGA- based compute nodes and has the capability of 6.4 Million 1D correlations per second for HSI endmember selection, classification and detection.

This paper presents a morphological polynomial approach to the presentation, shape decomposition, and object pattern recognition for machine vision. The polynomial morphological approach is a powerful method for image processing and pattern recognition. The algorithms of shape decomposition and pattern recognition are discussed. Polynomial approach can be implemented in a parallel processing machine and it can be used to develop a standard algebra-based programming language for image processing and pattern recognition.

The aim of this work is to explore the applicability of a relatively new snakes formulation called geometric snakes, for robust contour segmentation in radar images. In particular, we are looking for clear experimental indicators regarding the usefulness of such tool for radar imagery. In this work, we mainly concentrate on various contour segmentation problems in airborne and spaceborne SAR images (swatch and inverse mode). As an example, we study the segmentation of coastlines and ship targets. We observe that the dynamical and adaptive properties of geometric contours is better suited to determine the morphological properties of the contours. For high-resolution radar images of ships, the underlying motivation is that these properties could help providing robust extraction of ship structures for automatic ship classification.

The Metropolis Monte Carlo (MMC) Annealing is presented. In this approach to deconvolution two Monte Carlo Procedure (MCP) are run at the same time. In one the blurred data is used as a distribution function for selection of pixels. And the second MCP decides whether to place a grain in the true data (true input) or not. We show that this approach improves the annealing procedure drastically as compared to selection of pixels one at a time or from a flat distribution. The blurred data is obtained by convolving a 24 points input signal that has three peaks with a 21 points wide Gaussian impulse response function. The Mean Squared Error (MSE) is used to compare the two techniques. The MSE is calculated by comparing the reconstructed input signal with the true input signal. The MSE in reconstructed blurred data performed by MMC is also plotted vs. Monte Carlo move. Finally, the reconstructed input signal by MMC techniques is given at MSE of 39.

In this paper, we present a robust image filter that provides preservation of fine details and effective suppression of intensive multiplicative noise. The filter is based on the use of M (robust maximum likelihood)-estimators and R(rank)-estimators derived from the statistical theory of rank tests. At the first stage, to provide impulsive noise rejection, the introduced image filters uses the central pixel of the filtering window and the redescending M-estimators combined with the median estimator. At the second stage, to provide multiplicative noise suppression, a modified Sigma filter that implements the calculation scheme of a redescending M-estimator, is used. Visual and analytical analysis of simulation results shows that the proposed image filter has demonstrated fine detail preservation, good multiplicative noise suppression and impulsive noise removal.

The paper presents a new method for video compression based on the enhanced embedded zerotree wavelet (EZW) scheme. Recently, video codecs from the EZW family which use a 3D version of EZW or SPIHT algorithms showed better performance than the MPEG-2 compression algorithm. These algorithms have many advantages inherent for wavelet based schemes and EZW- like coders. Their most important advantages are good compression performance and scalability. However, they still allow improvement in several ways. First, as we recently showed, using Karhunen-Loeve (KL) transform instead of wavelet transform along the time axis improves compression ratio. Second, instead of the 3D EZW quantization scheme, we offer to use a convenient 2D quantization for every decorrelated frame adding one symbol `Strong Zero Tree', which means that every frame from a chosen set has a zero tree in the same location. The suggested compression algorithm based on KL transform, wavelet transform, and a new quantization scheme with strong zerotrees is free from some drawbacks of the plain 3D EZW codec. The presented codec shows 1 - 6 dB better results compared to the MPEG-2 compression algorithm on video sequences with small and medium motion.