JPL is developing a comprehensive Automatic Target Recognition (ATR) system that consists of an innovative anomaly detection preprocessing module and an automatic training target recognition module. The anomaly detection module is trained with an imaging data feature retrieved from an imaging sensor suite that represents the states of the normalcy model. The normalcy model is then trained from a self-organizing learning system over a period of time and fed into the anomaly detection module for scene anomaly monitoring and detection. The "abnormal" event detection will be sent to a human operator for further investigation responses. The target recognition will be continuously updated with the "normal' input sensor data. The combination of the anomaly detection preprocessing module to the re-trainable target recognition processor will result in a dynamic ATR system that is capable of automatic detection of anomaly event and provide an early warning to a human operator for in-time warning and response.

Detection and tracking of illegally parked vehicles are usually considered as crucial steps in the development of a video-surveillance based traffic-management system. The major challenge in this task lies in making the tracking phase illumination-change tolerant. The paper presents a two-stage process to detect vehicles parked illegally and monitor these in subsequent frames. Chromaticity and brightness distortion estimates are used in the first stage to segment the foreground objects from the remainder of the scene. The process then locks onto all stationary 'vehicle'-size patches, parts of which overlap with the regions of interest marked interactively a priori. The second stage of the process is applied subsequently to track all the illegally parked vehicles detected during the first stage. All the locked patches are filtered using a difference-of-Gaussian (DoG) filter operated at three different scales to capture a broad range of information. In succeeding frames patches at the same locations are similarly DoG filtered at the three different scales and the results matched with the corresponding ones initially generated. A combined score based on correlation estimates is used to track and confirm the existence of the illegally parked vehicles. Use of the DoG filter helps in extracting edge based features of the patches thus making the tracking process broadly illumination-invariant. The two-stage approach has been tested on the United Kingdom Home Office iLIDS dataset with encouraging results.

Electrocardiography is a diagnostic procedure for the detection and diagnosis of heart abnormalities. The electrocardiogram (ECG) signal contains important information that is utilized by physicians for the diagnosis and analysis of heart diseases. So good quality ECG signal plays a vital role for the interpretation and identification of pathological, anatomical and physiological aspects of the whole cardiac muscle. However, the ECG signals are corrupted by noise which severely limit the utility of the recorded ECG signal for medical evaluation. The most common noise presents in the ECG signal is the high frequency noise caused by the forces acting on the electrodes. In this paper, we propose a new ECG denoising method based on the empirical mode decomposition (EMD). The proposed method is able to enhance the ECG signal upon removing the noise with minimum signal distortion. Simulation is done on the MIT-BIH database to verify the efficacy of the proposed algorithm. Experiments show that the presented method offers very good results to remove noise from the ECG signal.

Jet Propulsion Lab and Vescent Photonics Inc. and are jointly developing an innovative ultracompact (volume < 10 cm³), ultra-low power (<10^-3 Watt-hours per measurement and zero power consumption when not measuring), completely non-mechanical Liquid Crystal Waveguide Fourier Transform Spectrometer (LCWFTS) that will be suitable for a variety of remote-platform, in-situ measurements. These devices are made possible by novel electro-evanescent waveguide architecture, enabling "monolithic chip-scale" Electro Optic-FTS (EO-FTS) sensors. The potential performance of these EO-FTS sensors include: i) a spectral range throughout 0.4-5 μm (25000 - 2000 cm^-1), ii) high-resolution (Δλ ≤ 0.1 nm), iii) high-speed (< 1 ms) measurements, and iv) rugged integrated optical construction. This performance potential enables the detection and quantification of a large number of different atmospheric gases simultaneously in the same air mass and the rugged construction will enable deployment on previously inaccessible platforms. The sensor construction is also amenable for analyzing aqueous samples on remote floating or submerged platforms. We will report a proof-of-principle prototype LCWFTS sensor that has been demonstrated in the near-IR (range of 1450-1700 nm) with a 5 nm resolution. This performance is in good agreement with theoretical models, which are being used to design and build the next generation LCWFTS devices.

JPL has developed an innovative electro-optic time delay line and utilize it to build a prototype proof-of-principle completely non-mechanical Electro-optic Hyperspectral Imaging Fourier Transform Spectrometer (EOHIFTS). Due to the use of the EO time delay line, the EOHIFTS is lightweight, broad spectral band, hyperspectral resolution that cannot be achieved simultaneously by any of the Imaging Fourier Transform Spectrometers (IFTS) developed to date. We will report the recent progress in the development of a feasibility breadboard and its feasibility demonstration.

A frequency domain implementation of the Optimal Trade-off Maximum Average Correlation Height (OT-MACH) filter has been optimized to classify target vehicles acquired from a Forward Looking Infra Red (FLIR) sensor. The clutter noise does not have a white spectrum and models employing the power spectral density of the background clutter require a predefined threshold. A method of automatically adjusting the noise model in the filter by using the input image statistical information has been introduced. Parameter surfaces for the remaining OT-MACH variables are calculated in order to determine optimal operating conditions for the view independent recognition of vehicles in highly cluttered FLIR imagery.

A space domain implementation of the Optimal Trade-off Maximum Average Correlation Height (OT-MACH) filter can not only be designed to be invariant to change in orientation of the target object but also to be spatially variant, i.e. the filter function becoming dependant on local clutter conditions within the image. Sequential location of the kernel in all regions of the image does, however, require excessive computational resources. An optimization technique is discussed in this paper which employs low-pass filtering to highlight the potential region of interests in the image and then restricts the movement of the kernel to these regions to allow target identification. The detection and subsequent identification capability of the two-stage process has been evaluated in highly cluttered backgrounds using both visible and thermal imagery and associated training data sets. A performance matrix comprised of peak-to-correlation energy (PCE) and peak-to-side lobe ratio (PSR) measurements of the correlation output has been calculated to allow the definition of a recognition criterion. A feasible hardware implementation for potential use in a security application using the proposed two-stage process is also described in the paper.

There are applications that require detection of multiple features which remain consistent in shape locally, but may change position with respect to one another globally. We refer to these feature sets as multi-feature constellations. We introduce a multi-level correlation filter design which uses composite feature detection filters, which on one level detect local features, and then on the next level detect constellations of these local feature responses. We demonstrate the constellation filter method with sign language recognition and fingerprint matching.

This paper proposes a new pattern recognition system employing optical joint transform correlation (JTC) technique which offers a great number of advantages over similar digital techniques, including very fast operation, simple architecture and capability of updating the reference image in real time. The proposed JTC technique incorporates a synthetic discriminant function (SDF) of the target image estimated from different training images to make the pattern recognition performance invariant to noise and distortion. It then involves four different phase-shifted versions of the same target SDF reference image, which are individually joint transform correlated with the given input scene. When the correlation signals are combined, it produces a single cross-correlation peak corresponding to each potential target present in the given input scene. The proposed technique also includes a fringe-adjusted filter to generate a delta-like correlation peak with high discrimination between the target and the background noise. The pattern recognition performance is further enhanced by incorporating the color information of the target objects in the proposed technique. The proposed technique is investigated using computer simulation where it shows efficient and successful target detection performance in different complex environments.

In this paper a novel method is proposed and demonstrated for automatic rotation angle measurement of a 2D object using a hybrid architecture, consisting of a 4f optical correlator with a binary phase only multiplexed matched filter and a single layer neural network. The hybrid set-up can be considered as a two-layer perceptron-like neural network; an optical correlator is the first layer and the standard single layer neural network is the second layer. The training scheme used to train the hybrid architecture is a combination of a Direct Binary Search algorithm, to train the optical correlator, and an Error Back Propagation algorithm, to train the neural network. The aim is to perform the major information processing by the optical correlator with a small additional processing by the neural network stage. This allows the system to be used for real-time applications as optics has the inherent ability to process information in a parallel manner at high speed. The neural network stage gives an extra dimension of freedom so that complicated tasks like automatic rotation angle measurement can be achieved. Results of both computer simulation and experimental set-up are presented for rotation angle measurement of an English alphabetic character as a 2D object. The experimental set-up consists of a real optical correlator using two spatial light modulators for both input and frequency plane representations and a PC based model of a single layer network.

In order to image a large area with a required resolution, a traditional camera would have to scan a smaller field-of-view until the entire area of interest is covered, thus losing persistence. Using a large sensor would result in high bandwidth data streams along with expensive and heavy equipment. Ideally, one would like to sense (or measure) a large number of pixels with a very limited set of measurements. In such a scenario the theory of compressive sensing may be put to use. A single sensor compressive imager for the wide area surveillance problem has been postulated and shown to be effective in detecting moving targets in a wide area. In this paper we look at the compressive imaging problem by assuming we have multiple cameras at our disposal. We show that we can get significant benefit in image reconstruction from multiple cameras measuring overlapped fields-of-view without any intra-camera communications and under significant transmission bandwidth constraints. We also show analysis and experiments which suggest that we can register these multiple cameras given only the random projective measurements from each camera.

We present a multi-stage automatic target recognition (ATR) system using a kernel-based PCA (kPCA) for nonlinear feature extraction. The kPCA method uses a nonlinear kernel function to map data onto a higher dimensional space and then performs the PCA in the feature space. An algorithm for inserting kernel PCA into the existing ATR system was designed and various types of kernels were tested and optimized on several testing image sets such as video images of boats in choppy waves or approaching helicopters. We discuss the performance comparisons and trade-offs in using kPCA for ATR operations. kPCA generally outperforms normal PCA in classification accuracy and free-response receiver operating characteristics (FROC).

The LPED (local polar edge detection) method is a newly developed 2D image processing method that automatically utilizes the center-of-mass polar coordinate to represent, in a unique way by a 36-dimension analog vector, the boundary of each object embedded in a picture frame. This 36D vector is the object ID for the particular object it represents. This ID vector is independent of the position of the object and independent of the orientation of the object, but it is a characteristic property from object to object. The background noises are automatically filtered out if the background objects are much smaller and much more randomly distributed than the objects of interest. This concise ID vector will not only identify the object precisely in a large picture frame where multiple-shaped objects lie, it will also track the object automatically when the object moves and it will record the data of movement periodically. I.e., it can measure automatically the distance of movement, the angular change of object-orientation, and the new locations of the central of mass of the moving object between successive sampling time intervals. In other words, it can automatically predict the near future movement of the tracked object. The applications of this novel image processing technique, to name a few, may be (1) automatic satellite-tracking and targeting of ground moving vehicles, (2) robotic identification of surrounding environment by some shape selected scenic part in the environment (e.g., the cross-section of an underground tunnel) with self guidance for the robot to go along a desired path through the whole tunnel without hitting the tunnel wall. This paper describes the principle of LPED and some extensive experimental results, regarding the application (1) described above, by utilizing a real-time soft-ware program designed by the author.

We present an innovative way to autonomously classify LiDAR points into bare earth, building, vegetation, and other categories. One desirable product of LiDAR data is the automatic classification of the points in the scene. Our algorithm automatically classifies scene points using Compressed Sensing Methods via Orthogonal Matching Pursuit algorithms utilizing a generalized K-Means clustering algorithm to extract buildings and foliage from a Digital Surface Models (DSM). This technology reduces manual editing while being cost effective for large scale automated global scene modeling. Quantitative analyses are provided using Receiver Operating Characteristics (ROC) curves to show Probability of Detection and False Alarm of buildings vs. vegetation classification. Histograms are shown with sample size metrics. Our inpainting algorithms then fill the voids where buildings and vegetation were removed, utilizing Computational Fluid Dynamics (CFD) techniques and Partial Differential Equations (PDE) to create an accurate Digital Terrain Model (DTM) [6]. Inpainting preserves building height contour consistency and edge sharpness of identified inpainted regions. Qualitative results illustrate other benefits such as Terrain Inpainting's unique ability to minimize or eliminate undesirable terrain data artifacts.

In this paper, we demonstrate optical correlation via dynamic range compression in two-beam coupling using thin-film organic materials. In contrast to the first demonstration, in which it was not possible to demonstrate correlation with complicated input, here we demonstrate correlation with extremely challenging cases involving finger prints, images in clutter, and SAR images. Our correlation results outperform many correlation results, including ones based on optimal filters.

Imaging in atmospheric turbulence and target recognition in cluttered environments have been research topics for many years. Currently, there are some well-established techniques for image restoration and recognition; however, if the atmospheric turbulence becomes a severe scattering medium and the surrounding environment is very cluttered, most conventional methods, such as inverse filtering and Wiener filtering, will be inadequate for correcting and recognizing the captured images. In this paper, we experimentally demonstrate nonlinear dynamic range compression techniques for image restoration and correlation via two-beam coupling and four wave mixing in organic photorefractive films.

A robust human intrusion detection technique using hue-saturation histograms is presented in this paper. Initially a region of interest (ROI) is manually identified in the scene viewed by a single fixed CCTV camera. All objects in the ROI are automatically demarcated from the background using brightness and chromaticity distortion parameters. The segmented objects are then tracked using correlation between hue-saturation based bivariate distributions. The technique has been applied on all the 'Sterile Zone' sequences of the United Kingdom Home Office iLIDS dataset and its performance is evaluated with over 70% positive results.

Fingerprint recognition is one of the first techniques used for automatically identifying people and today it is still one of the most popular and effective biometric techniques. With this increase in fingerprint biometric uses, issues related to accuracy, security and processing time are major challenges facing the fingerprint recognition systems. Previous work has shown that polarization enhancementencoding of fingerprint patterns increase the accuracy and security of fingerprint systems without burdening the processing time. This is mainly due to the fact that polarization enhancementencoding is inherently a hardware process and does not have detrimental time delay effect on the overall process. Unpolarized images, however, posses a high visual contrast and when fused (without digital enhancement) properly with polarized ones, is shown to increase the recognition accuracy and security of the biometric system without any significant processing time delay.

The problem of optical character recognition (OCR) of handwritten Arabic has not received a satisfactory solution yet. In this paper, an Arabic OCR algorithm is developed based on Hidden Markov Models (HMMs) combined with the Viterbi algorithm, which results in an improved and more robust recognition of characters at the sub-word level. Integrating the HMMs represents another step of the overall OCR trends being currently researched in the literature. The proposed approach exploits the structure of characters in the Arabic language in addition to their extracted features to achieve improved recognition rates. Useful statistical information of the Arabic language is initially extracted and then used to estimate the probabilistic parameters of the mathematical HMM. A new custom implementation of the HMM is developed in this study, where the transition matrix is built based on the collected large corpus, and the emission matrix is built based on the results obtained via the extracted character features. The recognition process is triggered using the Viterbi algorithm which employs the most probable sequence of sub-words. The model was implemented to recognize the sub-word unit of Arabic text raising the recognition rate from being linked to the worst recognition rate for any character to the overall structure of the Arabic language. Numerical results show that there is a potentially large recognition improvement by using the proposed algorithms.

An inexpensive, non-intrusive, vision-based, active fatigue monitoring system is presented. The system employs a single consumer webcam that is modified to operate in the near-IR range. An active IR LED system is developed to facilitate the quick localization of the eye pupils. Imaging software tracks the eye features by analyzing intensity areas and their changes in the vicinity of localization. To quantify the level of fatigue the algorithm measures the opening of the eyelid, PERCLOS. The software developed runs on the workstation and is designed to draw limited computational power, so as to not interfere with the user task. To overcome low-frame rate and improve real-time monitoring, a two-phase detection and tacking algorithm is implemented. The results presented show that the system successfully monitors the level of fatigue at a low rate of 8 fps. The system is well suited to monitor users in command centers, flight control centers, airport traffic dispatchers, military operation and command centers, etc., but the work can be extended to wearable devices and other environments.

Image acquired by an electro-optical system is spatially prefiltered by the optics, than sampled by the detector and finally postfiltered digitally during image processing. A method is presented for maximizing system performance by derivation of an optimal Whitening Matched post-Filter (WMF). The derivation of the WMF is based on combining clutter aliasing, misregistration, detector noise and target aliasing into a unified 'colored' noise. Further system optimization is achieved by modification of the optical point spread function (prefilter) so 'sampling balanced' configuration is achieved. In this configuration undersampling noise mechanisms (Aliasing and misregistration) and oversampling noise mechanisms (detector noise) are balanced to contribute equal magnitude. The performance of sampling-balanced system having WMF is compared to conventional systems for point target detection. It is shown that this system is the optimal and most robust in case of variations in scenario parameters.

A novel bag-of-visual-words algorithm is presented with two extensions compared to its classical version: exploiting scale information and weighting visual words. The scale information that is already extracted with SIFT detector is included as an additional element to the SIFT key-point descriptor, while the visual words are weighted during histogram assignment proportional to their importance which is measured by the ratio of their occurrences in the object to the occurrences in the background. The algorithm is tested for different geo-spatial object classes and the performance of the classical bag-of-visual-words algorithm is compared against the classical approach. Based on these results, a significant improvement is observed in terms of detection performance.

Image registration plays a vital role in many real time imaging applications. Registering the images in a precise manner is a challenging problem. In this paper, we focus on improving image registration error computation using the projection onto convex sets (POCS) techniques which improves the sub-pixel accuracy in the images leading to better estimates for the registration error. This can be used in turn to improve the registration itself. The results obtained from the proposed technique match well with the ground truth which validates the accuracy of this technique. Furthermore, the proposed technique shows better performance compared to existing methods.

In this paper, we present a new algorithm for target detection using hyperspectral imagery. The proposed algorithm is inspired by the outstanding performance of nonlinear RX-algorithm and the robustness of the stochastic expectation maximization (SEM) algorithm. The traditional technique of using SEM algorithm for target detection in hyperspectral imagery is associated with dimensionality reduction of the input data using binning or principal components analysis (PCA) algorithm. Although, the data reduction of the input data is enforced to reduce the computational burden on SEM algorithm, but it affects the results of target detection, especially the challenging one, due to not using the entire information of the potential targets. To facilitate detection of the target by using the entire targets information and simultaneously reducing the computational burden on SEM algorithm, we propose a new scheme for data reduction based on using Kernels. Kernel-based input data reduction is a nonlinear filtering technique in which the input data are mapped to the feature space where most of the background data is filtered using an easily selected threshold. Then, Gaussian mixture model is generated for the reduced input-data and SEM algorithm is employed to estimate the model parameters and to classify that input data. Finally, we allocated the target's class and isolated the target pixels. The proposed scheme for fusion the kernel with SEM algorithm has been tested using real life hyperspectral imagery and the results show superior performance compared to alternate algorithms.

The NarTest fluorescent technique is aimed at the detection of analyte of interest in street samples by recognition of its specific spectral patterns in 3-dimentional Spectral Fluorescent Signatures (SFS) measured with NTX2000 analyzer without chromatographic or other separation of controlled substances from a mixture with cutting agents. The illicit drugs have their own characteristic SFS features which can be used for detection and identification of narcotics, however typical street sample consists of a mixture with cutting agents: adulterants and diluents. Many of them interfere the spectral shape of SFS. The expert system based on Artificial Neural Networks (ANNs) has been developed and applied for such pattern recognition in SFS of street samples of illicit drugs.

Detection probability is an important index to represent and estimate target viability, which provides basis for target recognition and decision-making. But it will expend a mass of time and manpower to obtain detection probability in reality. At the same time, due to the different interpretation of personnel practice knowledge and experience, a great difference will often exist in the datum obtained. By means of studying the relationship between image features and perception quantity based on psychology experiments, the probability model has been established, in which the process is as following.Firstly, four image features have been extracted and quantified, which affect directly detection. Four feature similarity degrees between target and background were defined. Secondly, the relationship between single image feature similarity degree and perception quantity was set up based on psychological principle, and psychological experiments of target interpretation were designed which includes about five hundred people for interpretation and two hundred images. In order to reduce image features correlativity, a lot of artificial synthesis images have been made which include images with single brightness feature difference, images with single chromaticity feature difference, images with single texture feature difference and images with single shape feature difference. By analyzing and fitting a mass of experiments datum, the model quantitys have been determined. Finally, by applying statistical decision theory and experimental results, the relationship between perception quantity with target detection probability has been found. With the verification of a great deal of target interpretation in practice, the target detection probability can be obtained by the model quickly and objectively.

The paper considers neural net models and training and recognizing algorithms with base neurobiologic operations: p-step autoequivalence and non-equivalenc The Modified equivalently models (MEMs) of multiport neural net associative memory (MNNAM) are offered with double adaptive - equivalently weighing (DAEW) for recognition of 2D-patterns (images). It is shown, the computing process in MNNAM under using the proposed MEMs, is reduced to two-step and multi-step algorithms and step-by-step matrix-matrix (tensor-tensor) procedures. The given results of computer simulations confirmed the perspective of such models. Besides the result was received when MNNAM capacity on base of MEMs exceeded the amount of neurons.

In this paper, double-random phase-encoding based image hiding method was employed to encrypt and hide text. The ASCII codes of the secret text information was denoted as binary, and then transformed to a 2-dimensional array in the form of an image. Each element in the transformed array has a value between 0 and 255, where the highest 2 bits or the highest 4 bits were stored with the binary bits of the text information, while the lower bits were filled with binary bits. Then, the double-random phase-encoding method was used to encode the transformed array, and the encoded array was hidden into an expanded cover image to achieve text information hiding. Experimental results show that the secret text can be recovered accurately with the ratio of 100% and 99.89% by storing the binary bits of the text information to the highest 2 bits and the highest 4 bits of the transformed array, respectively. By employing the optical information processing method, the proposed method can improve the security of text information transmission, while keeping high hiding capacity.