We present an overview of a method using Independent Component Analysis (ICA) and 3D Integral Imaging (II) technique to recognize 3D objects at different orientations. This method has been successfully applied to the recognition and classification of 3D scenes.

We have continued to develop the Grayscale Optical Correlator (GOC) system and have explored a variety of automatic target recognition (ATR) applications to take advantage of the inherent performance advantages of the GOC vast parallelism and high-speed [1-4]. Recently, we have added 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. In this paper, we will discuss recent advancements in both the ATR processing algorithm development as well as an innovative breakthrough in designing new GOC hardware system architecture. First, we will briefly overview recent advances in our GOC and NN processor algorithm development. We will then present a new architecture that can lead to the mass production of a new generation of high-performance, low-cost Grayscale Optical Correlator. This new GOC architecture relies on the utilization of the maturing Digital Light Processor (DLP) as both the input and the filter Spatial Light Modulator (SLM). Detailed system description and performance analysis will also be reported.

We propose a novel space domain volume holographic correlator system. One of the limitations of conventional correlators is the bandwidth limits imposed by updating the filter and the readout speed of the CCD. The volume holographic correlator overcomes these by storing a large number of filters that can be interrogated simultaneously. By using angle multiplexing, the match can be read out onto a high speed linear array of sensors. A scanning window can be used to implement shift invariance, thus, making the system operate like a space domain correlator. The space domain correlation method offers an advantage over the frequency domain correlator in that the correlation filter no longer has shift invariance imposed on it since the kernel can be modified depending on its position. This maybe used for normalising the kernel or imposing some non-linearity in an attempt to improve performance. However, one of the key advantages of the frequency domain method is lost using this technique, namely the speed of the computation. A large kernel space-domain correlation, performed on a computer, will be very slow compared to what is achievable using a 4f optical correlator. We propose a method of implementing this using the scanning holographic memory based correlator.

In most object recognition applications, the object is present with different distortions (e.g. aspect view and scale) and its location is unknown. Our objective is to develop higher-order classifiers that can be applied (efficiently and fast) for different locations of the object over the test input. A type of classifier, the distortion-invariant filter (DIF), is attractive for fast object recognition, since it can be applied for different shifts using the fast Fourier transform (FFT); a single DIF handles different object distortions, e.g. all aspect views and some range of scale. In our prior work (Proc. SPIE 7252- 02), we combined DIFs and the kernel technique to form higher-order "kernel DIFs." In this paper, we present new test results with these kernel DIFs; we emphasize kernel versions of the synthetic discriminant function (SDF) filter, since we recall that they are the most efficient to use. We include new insight into the difference between vector-based and pixel-based kernels. We also present more test results with our recently introduced (Proc. SPIE 6977-03) combination of minimum noise and correlation energy (MINACE) filter preprocessing and kernel SDF filters (these form "preprocessed kernel SDF filters"); in our new work, we consider whether automated selection of the Minace-preprocessing parameter improves filter performance. We consider the classification of different pairs of true-class CAD (computer-aided design) infrared (IR) objects and the rejection of unseen problematic (blob) real IR clutter and unseen confuser-class CAD IR objects with full 360° aspect-view distortions and with different ranges of scale distortions present. We present new test results with more and different confuser-class objects and for both polynomial and Gaussian kernel SDF filters. We also include new test results at farther ranges than before; these are emphasized in this paper.

O-rings are regularly replaced in aircraft and if they are not replaced or if they are installed improperly, they can result in catastrophic failure of the aircraft. It is critical that the o-rings be packaged correctly to avoid mistakes made by technicians during routine maintenance. For this reason, fines may be imposed on the o-ring manufacturer if the o-rings are packaged incorrectly. That is, a single o-ring must be packaged and labeled properly. No o-rings or more than one o-ring per package is not acceptable. We present an industrial inspection system based on real-time composite correlation filtering that has successfully solved this problem in spite of opaque paper o-ring packages. We present the system design including the composite filter design.

An automatic Optimum Trade-off Maximum Average Correlation Height (OT-MACH) filter generator for use in a gray-scale optical correlator (GOC) has been developed for improved target detection at JPL. While the OT-MACH filter has been shown to be an optimal filter for target detection, actually solving for the optimum is too computationally intensive for multiple targets. Instead, an adaptive step gradient descent method was tested to iteratively optimize the three OT-MACH parameters, α, β, and γ. The feedback for the gradient descent method was a composite of the performance measures, correlation peak height and peak to side lobe ratio. The automated method generated and tested multiple filters in order to approach the optimal filter quicker and more reliably than the current manual method. Initial usage and testing has shown preliminary success at finding an approximation of the optimal filter, in terms of α, β, γ values. This corresponded to a substantial improvement in detection performance where the true positive rate increased for the same average false positives per image.

An optoelectronic neural network based detection technique is proposed for multi-class distortion-invariant pattern recognition. The neural network is utilized in the training stage for a sequence of multi-class binary and gray level images for supervised learning using shifted phase-encoded joint transform correlator with fringe adjusted filter in the hidden layer to create composite images that are invariant to distortion. Simulation results show that the proposed technique is efficient in recognizing targets in variable environmental conditions.

Fourier correlators perform space-invariant linear filtering on all input points, so they can identify and locate patterns in parallel. Each output point is a weighted sum of components of the Fourier transform of the input, so the discriminants used are inherently linear. As most practical problems are not linearly discriminable, that causes a problem. This paper describes a quite general solution involving nonlinear combining of nonlinearly processed outputs from multiple Fourier masks. The design of the masks and nonlinearities allows very powerful nonlinear discrimination that preserves the space-invariant feature that makes Fourier correlators attractive. Given a set of target-class images, henceforth referred to as the training set or trainers, the algorithm developed herein computes an ordered set of classifier filters - Generalized Matched Filters (GMFs) threshold values. An unlabeled image is applied to the classifier filter set, hereafter referred to as super-generalized matched filter (SGMF). If the peak response of any of the classifier filters (GMFs) to the unlabeled test image exceeds the threshold level the decision is made in favor of labeling the image as target-class otherwise it is labeled non-target- class.

Linear methods of restoration of input scene's images in optical-digital correlators are described. Relatively low signal to noise ratio of a camera's photo sensor and extensional PSF's size are special features of considered optical-digital correlator. RAW-files of real correlation signals obtained by digital photo sensor were used for input scene's images restoration. It is shown that modified evolution method, which employs regularization by Tikhonov, is better among linear deconvolution methods. As a regularization term, an inverse signal to noise ratio as a function of spatial frequencies was used. For additional improvement of restoration's quality, noise analysis of boundary areas of the image to be reconstructed was performed. Experimental results on digital restoration of input scene's images are presented.

The results of realization of linear phase coefficient composite (LPCC) filters in 4-f correlator are presented in this paper. LPCC filters application allows achieving invariance of correlation peak in the presence of geometric distortions of contour objects. Invariant correlation filters were synthesized as amplitude holograms with continuous transparency and realized as binary amplitude holograms. The results of correlator signals modeling for the cases with different numbers of hologram counts which used to represent grayscale level gradation are presented.

We describe an automated remote cyclone detection and tracking approach using heterogeneous data from multiple satellites. Single Earth orbiting satellite has been used in the past to detect and track events such as cyclones but suffer from major drawbacks due to limited spatio-temporal coverage. Our novel approach addresses the challenges in using heterogeneous data from multiple data sources for knowledge discovery, tracking and mining of cyclones. Moreover, it offers better detection performance and spatio-temporal resolutions. Our solution is sufficiently powerful that it generalizes to multiple sensor measurement modalities. Our approach consists of: (i) feature extraction from each sensor measurement, (ii) an ensemble classifier for cyclone detection, and (iii) knowledge sharing between the different remote sensor measurements. Our extensive experimental results demonstrate (i) the superior performance of our cyclone detector compared to previous work on preprocessed historical data, (ii) stable performance of our cyclone detector when it is applied on different geographical regions (Western Pacific Ocean and the North Atlantic Ocean), (iii) meaningful knowledge derived from the cyclone detector output, and (iv) the performance quality of our automated cyclone detection and tracking solution closely match the cyclone best track information from the National Hurricane Center.

In this paper, we review the recent trends and advancements in decision fusion based target tracking in FLIR image sequences. In particular, we discuss four target tracking algorithms and two data fusion algorithms that have been used for single/multiple target detection and tracking purposes. Each tracking algorithm utilizes various properties of targets and image frames of a given sequence. The data fusion algorithms employ complementary features of two or more of the above mentioned algorithms. The data fusion technique has been found to yield better performance compared to the alternate algorithms as shown by the test results obtained using real life FLIR image sequences.

Much attention has been given to the visual tracking problem due to its obvious uses in military surveillance. However, visual tracking is complicated by the presence of motion of the observer in addition to the target motion, especially when the image changes caused by the observer motion are large compared to those caused by the target motion. A method is presented for estimating the motion of the observer based on image registration techniques and Kalman filtering. With the effects of the observer motion removed, an additional phase is implemented to track individual targets. This method is demonstrated on an image stream from a buoy-mounted or periscope-mounted camera, where large inter-frame displacements are present due to the wave action on the camera. This system has been shown to be effective at tracking and predicting the global position of a planar vehicle (boat) being observed from a single, out-of-plane camera. Finally, the tracking system has been extended to a multi-target scenario.

Our group has developed several methods for acquiring 3-dimensional objects in motion which include facial expressions. For this to be practical we need to identify and track various features contained in facial expressions. To accomplish this we introduce a set of feature based trackers and propose strategies for combining them together to form meshes. We present our strategy in the context of swarm theory where the elements of the swarm are the feature trackers and the communication structure of the swarm is essentially a spatial mesh. We demonstrate the concepts with examples of facial feature tracking.

Several approaches have been presented to enhance the recognition capabilities of certain patterns. This is particularly important in applications involve recognizing the identity of some individual accurately, especially in critical applications such as border entry, access to secured buildings, and in financial transactions, among other things. In this paper, we present an approach to improve the accuracy of a biometric recognition system using two fused patterns (two fingers) of same classifier (individual). In this method, two or more fingers of the same classifier are fused together to form a unique fingerprint pattern to be used in the recognition process. Techniques related to pattern and decision fusions are tested. In particular, the logical AND operator used in the decision fusion algorithm has resulted in a higher level of accuracy of recognition.

We propose an improvement on the existing super-resolution technique that produces high resolution video from low quality low resolution video. Our method has two steps: (1) motion registration and (2) regularization using back-projection. Sub-pixel motion parameters are estimated for a group of 16 low resolution frames with reference to the next frame and these are used to position the low resolution pixels on high resolution grid. A gradient based technique is used to register the frames at the sub-pixel level. Once we get the high resolution grid, we use an improved state-of-the-art regularization technique where the image is iteratively modified by applying back-projection to get a sharp and undistorted image. This technique is based on bilateral prior and deals with different data and noise models. This computationally inexpensive method is robust to errors in motion/blur estimation and results in images with sharp edges. The proposed system is faster than the existing ones as the post-processing steps involved only simple filtering. The results show the proposed method gives high quality and high resolution videos and minimizes effects due to camera jerks. This technique can easily be ported to hardware and can be developed into a product.

Nonlinear information processing via two-beam coupling using thin-film organic photorefractive material is demonstrated. The organic material is found to possess superior response time and resolution compared to photorefractive bulk material. The possibility of designing dynamic range compression deconvolution for restoring blurred images embedded in a noisy environment is also demonstrated.

A multi-stage automated target recognition (ATR) system has been designed to perform computer vision tasks with adequate proficiency in mimicking human vision. The system is able to detect, identify, and track targets of interest. Potential regions of interest (ROIs) are first identified by the detection stage using an Optimum Trade-off Maximum Average Correlation Height (OT-MACH) filter combined with a wavelet transform. False positives are then eliminated by the verification stage using feature extraction methods in conjunction with neural networks. Feature extraction transforms the ROIs using filtering and binning algorithms to create feature vectors. A feed forward back propagation neural network (NN) is then trained to classify each feature vector and remove false positives. This paper discusses the test of the system performance and parameter optimizations process which adapts the system to various targets and datasets. The test results show that the system was successful in substantially reducing the false positive rate when tested on a sonar image dataset.

This paper describes a real-time system for detecting people in infrared video taken by a re-locatable camera tower suitable for border monitoring. Wind effects cause the camera to sway, so typical background modeling techniques prove difficult to apply. Instead, detection is performed using a supervised classifier over a set of seven Scale Invariant Image Moments. Blobs images are generated with a simple application of thresholding and dilation, yielding a set of possible targets. For each potential target, the Scale Invariant Moments are computed and classified as "Person" or "Non-Person." We present three methods for training the classifier: Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and a two-layer Neural Network (NN). We compare the accuracy for the three methods. Results are presented for sample videos, showing acceptable accuracy while maintaining real time throughput. The key advantages of this method are real-time performance and tolerance of random ego motion.

We have developed a very efficient image preprocessing scheme that allows the main features (the boundary lines and the corner points) of some size-selected objects in a colored picture be singled out and be put in a much smaller bmp file. Then a recently corrected preprocessing program written in Visual Basic 6 can be applied directly to this file. We can then obtain a very compact analog data set to describe all the boundary lines contained in this file. This set of analog data can be used independently and efficiently to reconstruct all the main features of the selected objects. Before the reconstruction, we can also use an analytical filter to select some particular features according to, for example, the length of the boundary lines, the curvature of the boundary lines, the starting and ending points of the boundary lines, etc. These analytically filtered lines or curves can then be re-displayed in a different window and be used in some special-purpose object identification system that may accurately identify certain particular object from a very complex noisy background. Preliminary live experiments will be demonstrated.

θThe window unit in the design of the complex logarithmic r-θ mapping for hybrid optical neural network filter can allow multiple objects of the same class to be detected within the input image. Additionally, the architecture of the neural network unit of the complex logarithmic r-θ mapping for hybrid optical neural network filter becomes attractive for accommodating the recognition of multiple objects of different classes within the input image by modifying the output layer of the unit. We test the overall filter for multiple objects of the same and of different classes' recognition within cluttered input images and video sequences of cluttered scenes. Logarithmic r-θ mapping for hybrid optical neural network filter is shown to exhibit with a single pass over the input data simultaneously in-plane rotation, out-of-plane rotation, scale, log r-θ map translation and shift invariance, and good clutter tolerance by recognizing correctly the different objects within the cluttered scenes. We record in our results additional extracted information from the cluttered scenes about the objects' relative position, scale and in-plane rotation.

It is proposed an image complexity matrix (IC), based on the analysis of the Fourier spectrum of the input image. It is described the method of IC calculation. It was carried out the determination of the necessary number of the image's pixels dependent on the image complexity. It is presented the optical - electronic processor for IC determination. It is analyzed the structure of the optical electronic computer system for pattern and target recognition.

Being a popular choice among the biometric features, the fingerprint has long been used for identification and verification purposes by security agencies. In this paper, additional images are included in the input joint images in a joint transform correlator (JTC) to achieve a fast real-time fingerprint verification. In the proposed scheme, in addition to the reference and the target image, the input joint image has a complementary reference image and a complementary target image. At the correlation output of the JTC, the cross-correlation peak value between the reference and the complementary target image and the cross-correlation peak value between the complementary reference and the target images are used as the criteria to perform the recognition of the target in the input scene. It will be shown that these two cross-correlation peak values will be zero if and only if the input target matches the reference image. The scheme is employed to verify binary characters and binarized fingerprint images.

Most of the algorithm proposed for face recognition involve considerable amount of computations and hence they cannot be used on devices constrained with limited memory. In this paper, we propose a novel solution for efficient face recognition for systems which are charactered with small memory devices and demand fast performance. By dividing the face images into components, the scheme finds the discriminant phases of the Fourier transform of these components of face images. The discriminant phases are found by using linear discriminant analysis (LDA) to obtain a system that matches the criteria imposed by devices of limited memory and requiring fast recognition. Note that since the effects of face variation within an image are reduced when the image is divided into components, then the performance of the system is enhanced. Compared to spatial domain analysis, it will be shown that the use of component in the frequency domain produces better recognition performance. A thorough study and comprehensive experiments relating time consumption and computational complexity versus system performance are presented. The proposed technique increases the face recognition rate and at the same time achieves substantial saving in the computational time, when compared to other known systems. The experimental results reveal that a recognition rate of ≥99% is achieved, when applied to two independent and well known databases.

Reconstruction of missing parts or scratches of digital images is an important field used extensively in artwork restoration. This restoration can be done by using two approaches, image inpainting and texture synthesis. There are many techniques for the two pervious approaches that can carry out the process optimally and accurately. In this paper the advantages and disadvantages of most algorithms of the image inpainting approach are discussed. The modification to Oliveira inpainting model is introduced. This modification produces fast and good quality with one iteration without blur and removes large object with symmetric background.

In an infrared search and tracking (IRST) system, the clustering procedure which merges target pixels into one cluster requires larger computational load according to increasing clutters. In this paper, we propose a novel clustering method based on weighted sub-sampling to reduce clustering time and obtain suitable cluster in cluttered environment. A conventional sub-sampling method can reasonably reduce clustering time but cause large error, when obtaining cluster center. However, our proposed clustering method perform sub-sampling and assign specific weights which is the number of target pixels in sampling region to sub-sampled pixels to obtain suitable cluster center. After performing clustering procedure, the cluster center position is properly obtained using sampled pixels and their weights in the cluster. Therefore, our proposed method can not only reduce clustering time using a sub-sampling method, but also obtain proper cluster center using our proposed weights. To validate our proposed method, experimental results for several infrared and noise images are presented.

This work contains the results of the experiments on the restoration of the defective images proceeded in a matrix and a vector form with the help of the feed forward neural network. Sometimes it is convenient to represent an image as a vector rather than as a matrix. So the target of this work is to show experimentally what kind of input provides a better restoration, judging from the Euclid's distance of the output of a trained network. This work also shows the differences between processing different types of image presentation of the neuron network. Making a comparative analysis of a matrix and a vector form of presenting the images which are proceeded to a feed forward network allows stating some specific characteristics of a network. These characteristics include the optimal architecture of a network, the number of layers, the number of neurons in each layer and the time of an image restoration. Taking into account the network's characteristics and the most important factor - the Euclid's distance, are drawn conclusions that concern what is the best way of representing images that we want to restore using a feed forward network.