This PDF file contains the front matter associated with SPIE Proceedings Volume 6574, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

This paper presents the status of our SAR automatic target recognition (ATR) work on the Moving and Stationary Target Acquisition and Recognition (MSTAR) public database using the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). In our previous work, we used the MSTAR public database benchmark three-class problem and demonstrated better results than all prior work. In this paper, we address classification (including variants) and object and clutter rejection tests on the more challenging MSTAR ten-class public database. The Minace algorithm is shown to generalize well to this larger classification problem. We use several filters per object, but fewer DIFs per object than prior work did. We use our autoMinace algorithm that automates selection of the Minace filter parameter c and selection of the training set images to be included in the filter. No confuser, clutter, or test set data are present in the training or the validation set. In tests, we do not assume that the test input's pose is known (as most prior work does), since pose estimation of SAR objects has a large margin of error. We also address tests with proper use of SAR pose estimates in MSTAR recognition and the use of multilook SAR data to improve performance.

Three-dimensional light structures can be created by modulating the spatial phase and polarization properties of an an expanded laser beam. A particularly promising technique is the Generalized Phase Contrast (GPC) method invented and patented at Risø National Laboratory. Based on the combination of programmable spatial light modulator devices and an advanced graphical user-interface the GPC method enables real-time, interactive and arbitrary control over the dynamics and geometry of synthesized light patterns. Recent experiments have shown that GPC-driven micro-manipulation provides a unique technology platform for fully user-guided assembly of a plurality of particles in a plane, control of particle stacking along the beam axis, manipulation of multiple hollow beads, and the organization of living cells into three-dimensional colloidal structures. Here we present GPC-based optical micromanipulation in a microfluidic system where trapping experiments are computer-automated and thereby capable of running with only limited supervision. The system is able to dynamically detect living yeast cells using a computer-interfaced CCD camera, and respond to this by instantly creating traps at positions of the spotted cells streaming at flow velocities that would be difficult for a human operator to handle.

In this paper we present an information sensing system which integrates sensing and processing resulting in the direct collection of data which is relevant to the exploitation application. Broadly, integrated sensing and processing (ISP) considers algorithms that are integrated with the collection of data. We demonstrate an ISP system which utilizes a near Infrared (NIR) Hadamard multiplexing imaging sensor. This prototype sensor incorporates a digital mirror array (DMA) device in order to realize a Hadamard multiplexed imaging system. Specific Hadamard codes can be sent to the sensor to realize inner products of the underlying scene rather than the scene itself. The developed ISP algorithm incorporates the exploitation tasks into the sensing by computing an ATR metric which directs the sensor to collect only the information relevant to the ATR problem. The result is a multiple resolution hyperspectral cube with full resolution where targets are present and less than full resolution where there are no targets. We demonstrate this algorithm fully integrated with the sensor and running in real time on a test case to demonstrate feasibility.

We have been developing Grayscale Optical Correlator (GOC) and exploring a variety of automatic target recognition (ATR) applications to take advantage of the inherent performance advantages of the GOC parallel processing, high-speed, vast parallelism and high-speed [1-4]. To date, we have built compact 512 x 512, 1000 fps GOC systems and tested/demonstrated for field ATR experiments. We have also worked with our industrial partners to develop a 1024 x 1024 Ferroelectric Spatial Light Modulator (FLCSLM) to meet the challenging applications demanding larger input scene Field-of-View (FOV) and higher resolution. In this paper, two major system issues that we have encountered during the development efforts for real-world applications will be discussed. These include: 1) SLM dynamic range limitations and 2) ATR performance for CAD/CAC, computer-aided detection & classification (CAD/CAC) applications. Our simulation study has shown that the current 8-bit dynamic range possessed by the FLCSLM is adequate for both the input image and the correlation filter encodings. We will also describe the addition of a neural network (NN) post-processor to greatly decrease the false positive detection rate while retaining the high positive detection rate obtained by the by the GOC. Experimental results demonstrating the high-performance of the fused GOC and NN processor will be provided.

Detection of rotationally distorted targets is a challenging task in pattern recognition applications. Recently, we proposed and implemented a wavelet-modified maximum average correlation height (MACH) filter for in-plane and out-of-plane rotation invariance in hybrid digital-optical correlator architecture. Use of wavelet transform improved the performance of the MACH filter by reducing the number of filters required for identifying a rotated target and enhancing the correlation peak intensity significantly. The output of a hybrid digital-optical correlator contains two autocorrelation peaks and a strong dc. To capture a desired single autocorrelation peak a chirp function with the wavelet-modified MACH filter was used. The influence of perturbations in hybrid digital-optical correlator has also been studied. Perturbations include, the effect of occlusion on input target, the effect of additive and multiplicative noise and their combined effect on input target, and the effect of occlusion of product function to be optically processed for obtaining the correlation outputs. The present paper reviews investigations on the hybrid digital-optical correlation scheme with special reference to the work carried out at the Photonics Division, IRDE Dehradun.

Pattern recognition in hyperspectral imagery often suffers from a number of limitations, which includes computation complexity, false alarms and missing targets. The major reason behind these problems is that the spectra obtained by hyperspectral sensors do not produce a deterministic signature, because the spectra observed from samples of the same material may vary due to variations in the material surface, atmospheric conditions and other related reasons. In addition, the presence of noise in the input scene may complicate the situation further. Therefore, the main objective of pattern recognition in hyperspectral imagery is to maximize the probability of detection and at the same time minimize the probability of generating false alarms. Though several detection algorithms have been proposed in the literature, but most of them are observed to be inefficient in meeting the objective requirement mentioned above. This paper presents a novel detection algorithm which is fast and simple in architecture. The algorithm involves a Gaussian filter to process the target signature as well as the unknown signature from the input scene. A post-processing step is also included after performing correlation to detect the target pixels. Computer simulation results show that the algorithm can successfully detect all the targets present in the input scene without any significant false alarm.

In most ATR applications, objects are not only present with thermal and aspect view angle variations, its size (range) also changes as the sensor approaches the target, and depression angle variations can exist. Therefore, it is important and realistic to know how to handle these variations. We apply our new SVRDM (support vector representation and discrimination machine) classifier to address these problems. The SVRDM classifier has good generalization (like the standard SVM does), and it has the added property of a good rejection ability. In other words, it not only gives very promising recognition results on the true target classes, it is also able to reject other unseen objects (referred to as confusers). We address the following variation issues: the scale range one SVRDM can recognize when trained on data at one or more ranges, the depression angle difference one SVRDM can recognize when trained on data at only one (or several) depression angles, and the number of aspect views needed to be included in the training set to handle recognition of targets with aspect variations, and the classification and rejection performance. Thus, our results are most unique and worthwhile but are not easily compared to prior work. Recognition and rejection test results are presented on both simulated and real infra-red (IR) data.

In many tracking applications, adapting the target appearance model over time can improve performance. This approach is most popular in high frame rate video applications where latent variables, related to the objects appearance (e.g., orientation and pose), vary slowly from one frame to the next. In these cases the appearance model and the tracking system are tightly integrated, and latent variables are often included as part of the tracking system's dynamic model. In this paper we describe our efforts to track cars in low frame rate data (1 frame / second), acquired from a highly unstable airborne platform. Due to the low frame rate, and poor image quality, the appearance of a particular vehicle varies greatly from one frame to the next. This leads us to a different problem: how can we build the best appearance model from all instances of a vehicle we have seen so far. The best appearance model should maximize the future performance of the tracking system, and maximize the chances of reacquiring the vehicle once it leaves the field of view. We propose an online feature selection approach to this problem and investigate the performance and computational trade-offs with a real-world dataset.

A method of recognising and tracking multiple solid objects in video sequences despite any kind of perspective distortion is demonstrated. Moving objects are initially segmented from the scene using a background subtraction method to minimize the search area of the filter. A variation on the Maximum Average Correlation Height (MACH) filter is used to create invariance to orientation while giving high tolerance to background clutter and noise. A log r-θ mapping is employed to give invariance to in-plane rotation and scale by transforming rotation and scale variations of the target object into vertical and horizontal shifts. The MACH filter is trained on the log r-θ map of the target for a range of orientations and applied sequentially over the regions of movement in successive video frames to test for target objects. A Kalman filter is employed to continuously track the target objects over successive frames, which has enabled the system to track multiple targets despite temporary occlusion or intersection.

A electro-optic (EO) and infrared (IR) automatic target recognition (ATR) system based on the minimum noise and correlation energy (MINACE) distortion invariant filter (DIF) is presented. The system uses exceptionally high resolution EO and IR data obtained from the Shared Reconnaissance Pod (SHARP). Excellent detection results are obtained. Furthermore, the selection of a key parameter - the MINACE filter parameter c - is fully automated using a training and validation set. We also present a set of correlation plane post processing methods to reduce false alarms and improve detection accuracies. The system is evaluated using multi-sensor imagery acquired using the SHARP sensor suite, the detection (P_D) and false alarm (P_FA) scores are presented for the problem of detecting aircrafts in the high resolution imagery. The scale and orientation of the targets are not assumed to be known, thus making the problem more realistic.

Hyperspectral imagery is used for a wide variety of applications, including target detection, tacking, agricultural monitoring and natural resources exploration. The main reason for using hyperspectral imagery is that these images reveal spectral information about the scene that are not available in a single band. Unfortunately, many factors such as sensor noise and atmospheric scattering degrade the spatial quality of these images. Recently, many algorithms are introduced in the literature to improve the resolution of hyperspectral images [7]. In this paper, we propose a new method to produce high resolution bands from low resolution bands that are strongly correlated to the corresponding high resolution panchromatic image. The proposed method is based on using the local correlation instead of using the global correlation to improve the estimated interpolation in order to construct the high resolution image. The utilization of local correlation significantly improved the resolution of high resolution images when compared to the corresponding results obtained using the traditional algorithms. The local correlation is implemented by using predefined small windows across the low resolution image. In addition, numerous experiments are conducted to investigate the effect of the chosen window size in the image quality. Experiments results obtained using real life hyperspectral imagery is presented to verify the effectiveness of the proposed algorithm.

In this paper the A-law/μ-law Dynamic Range Compression algorithm used in telecommunication systems is proposed for the first time for nonlinear Dynamic Range Compression image deconvolution. In the proposed setup, a joint image of the blurred input information and the blur impulse response are jointly Fourier-transformed via a lens to a CCD camera which acts as a square-law receiver. The CCD camera is responsible for mixing the Fourier transforms of the impulse response and the distorted image to compensate for the phase distortion and then the A-law/μ-law nonlinear transformation is responsible for enhancing both the high frequencies and the signal-to-noise ratio. The proposed technique is supported by computer simulation.

We present experimental results of training a neural network to perform chemical compound identification from a portable space-borne gas chromatographic mass spectrometer (GCMS). The GCMS data has distortion, peak overlap, and noise problems. A signal processing algorithm is first applied to the GCMS to detect the peaks and to clean the MS spectra. We design neural networks to be trained on a sub-set of chemicals that are closely related in the GC graph. Each sub-neural network then identifies the compounds within the sub-set. We design the training data using mostly NIST standard MS data. The NIST mass spectral data of multiple compounds are mixed to train the neural network to identify mixed species. Back-propagation learning algorithm is used to train the neural network. Good identification results have been obtained.

Any event deemed as being out-of-the-ordinary may be called an anomaly. Anomalies by virtue of their definition are events that occur spontaneously with no prior indication of their existence or appearance. Effects of anomalies are typically unknown until they actually occur, and their effects aggregate in time to show noticeable change from the original behavior. An evolved behavior would in general be very difficult to correct unless the anomalous event that caused such behavior can be detected early, and any consequence attributed to the specific anomaly. Substantial time and effort is required to back-track the cause for abnormal behavior and to recreate the event sequence leading to abnormal behavior. There is a critical need therefore to automatically detect anomalous behavior as and when they may occur, and to do so with the operator in the loop. Human-machine interaction results in better machine learning and a better decision-support mechanism. This is the fundamental concept of intelligent control where machine learning is enhanced by interaction with human operators, and vice versa. The paper discusses a revolutionary framework for the characterization, detection, identification, learning, and modeling of anomalous behavior in observed phenomena arising from a large class of unknown and uncertain dynamical systems.

A variety of autonomous sensory platforms for earth and planetary science require distributed sensing and control capability to be able to operate in dynamic environments where the occurrence and frequency of events could be few and far-between. We discuss a novel control methodology for autonomous monitoring of space habitats and in-situ ground and ocean-based heterogeneous wireless distributed sensor networks. Such sensor networks need to have a lifetime of months or even years, while being effective at detecting and reporting events in real-time before they pose a danger. This requires an on-line resource manager or controller to economize and adaptively control all resources such as energy, communication bandwidth, and sensor sampling frequency. We present an event based control optimization formulation of the resource management problem for sensor networks and discuss a method to adaptively change desired system performance of the sensor network in response to events. This functionality is critical in field-deployable sensor networks where continuous operation is expensive and system adaptation is critical for extended operation in the face of dynamic external events. We show results on various synthetic heterogeneous sensor networks where only partially accurate information about the sensing system is available and illustrate the efficacy of the control algorithm in handling such incorrect models with a negligible increase in transmission of the optimal control settings. We show that the run-time performance of the control algorithm scales well with increasing number of sensors.

This paper presents our IR automatic target recognition (ATR) work on the Comanche database using the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). The Comanche database contains real IR data of eight targets with aspect view and thermal state variations. We consider recognition of six of these targets and we consider rejecting two targets (confusers) and clutter. To handle the full 360° range of aspect view in Comanche data, we use a set of Minace filters for each object; each filter should recognize the object in some angular range. We use our autoMinace algorithm that uses a training and a validation set to select the Minace filter parameter c (which selects emphasis on recognition or discrimination) and to select the training set images to be included in the filter, so that the filter can achieve both good recognition and good confuser and clutter rejection performance. No confuser, clutter, or test set data are present in the training or the validation set. Use of the peak-to-correlation energy (PCE) ratio is found to perform better than the use of the correlation peak height metric. The use of circular versus linear correlations is addressed; circular correlations require less storage and fewer online computations and are thus preferable.

A system to detect vehicles (cars, trucks etc) in electro-optic (EO) and infrared (IR) imagery is presented. We present the use of the minimum noise and correlation (MINACE) distortion invariant filter (DIF) for this problem. The selection of the MINACE filter parameter c is automated using a training and validation set. A new set of correlation plane post processing methods that improve detection accuracies and reduce false alarms are presented. The system is tested on real life imagery of traffic in parking lots and roads obtained using a multi-sensor EO/IR platform.

Diffraction image correlator based on commercial digital SLR photo camera was reported earlier. The correlator was proposed for recognition of external scenes illuminated by quasimonochromatic spatially incoherent light. The correlator hardware consists of digital camera with plugged in optical correlation filter unit and control computer. The kinoform used as correlation filter is placed in a free space of the SLR camera body between the interchangeable camera lens and the swing mirror. On the other hand, this correlator can be considered as a hybrid optical-digital imaging system with wavefront coding. It allows not only to recognize objects in input scene but to restore, if needed, the whole image of input scene from correlation signals distribution registered by SLR camera sensor. Linear methods for image reconstruction in the correlator are discussed. The experimental setup of the correlator and experimental results on images recognition and input scenes restoration are presented.

Automatic adaptive tracking in real-time for target recognition provided autonomous control of a scale model electric truck. The two-wheel drive truck was modified as an autonomous rover test-bed for vision based guidance and navigation. Methods were implemented to monitor tracking error and ensure a safe, accurate arrival at the intended science target. Some methods are situation independent relying only on the confidence error of the target recognition algorithm. Other methods take advantage of the scenario of combined motion and tracking to filter out anomalies. In either case, only a single calibrated camera was needed for position estimation. Results from real-time autonomous driving tests on the JPL simulated Mars yard are presented. Recognition error was often situation dependent. For the rover case, the background was in motion and may be characterized to provide visual cues on rover travel such as rate, pitch, roll, and distance to objects of interest or hazards. Objects in the scene may be used as landmarks, or waypoints, for such estimations. As objects are approached, their scale increases and their orientation may change. In addition, particularly on rough terrain, these orientation and scale changes may be unpredictable. Feature extraction combined with the neural network algorithm was successful in providing visual odometry in the simulated Mars environment.

K-means clustering method has been employed in different applications of data analysis. This paper develops a target detection system using the k-means algorithm including a preprocessing step based on the Euclidean distance. The pre-processing step reduces the computational complexity of the k-means algorithm in case of hyperspectral imagery. After reducing the set of pixels in the background from the data by using the pre-processing step, k-means algorithm is employed to determine the clusters in rest of the image data cube. Having obtained the clustered data, the objects of interest can easily be detected using the known target signature. The proposed clustering algorithm is successfully applied to the real life hyperspectral data sets where the objects of interest can efficiently be detected. The proposed scheme effectively reduces the convergence time of the k-mean algorithm compared to that required by the traditional k-means algorithm.

A method of the recognition reliability estimation in the optical pattern recognition systems (OPRS) is described, based on of the similarity measures differences (SMD). It was theoretically justified and experimentally confirmed a hypothesis about the distribution law of the SMD. There were calculated the reliabilities of the correct objects recognition at single and coded correlation responses in OPRS of invariant and normalized images processing.

This paper deals with the problem of intellectual restoration of images. It is suggested to represent various objects and stages as objects of the first and second orders. Representation of dominant object as second order object reveals its new properties, that is an opportunity to control its own parameters. Complex representation of dominant object as second-class object of the first and second types allows to eliminate defects of its own image, as well as defects of image of subordinated object.