Abstract: This paper proposes a new speech recognition method using a tree-structured probability density function (PDF) to realize high speed HMM based speech recognition In order to reduce the likelihood calculation for a PDF set composed of the Gaussian PDFs for all mixture components, all states and all recognition units, it is coarsely done for the element PDF whose likelihood is not likely to be large The PDF set is expressed as a tree-structured form In the recognition process, the likelihood set is calculated by searching the tree; by calculating the likelihood from the cluster PDF at the node and traversing the nodes with the largest likelihood from the root Experimental results showed that the computation load was drastically reduced with little reduction in the recognition accuracy, in both speaker-independent and speaker-adaptive cases The algorithm was applied to a personal computer speech recognition software without using special hardware

Abstract: An apparatus for and method of classifying a pattern including performing a wavelet transform. The invention finds particular application in the field of automatic target recognition, in which the pattern appears in a return radar signal, and the classification is used to classify a target.

Abstract: We develop and extensively test a new algorithm for discriminating man-made objects from natural clutter in synthetic-aperture radar (SAR) imagery. The novel feature of our approach is its exploitation of the characteristically distinct variations in speckle pattern for imagery of man-made objects and of natural clutter, as image resolution is varied from coarse to fine. We treat these characteristics using stochastic framework, specifically tailored for multiresolution random processes and fields. Within the framework, we build a pair of multiscale models: one for SAR imagery of natural clutter and another for imagery of man-made objects. We then use these models to define a multiresolution discriminant as the likelihood ratio for distinguishing between the two image types, given a multiresolution sequence of images of a region of interest (ROI). We incorporate this likelihood ratio into an existing, established discriminator that was developed at Lincoln Laboratory (LL) as part of a complete system for automatic target recognition (ATR). To classify a given ROI, we merge the information provided by our likelihood ratio with the measured values of a small number of size and brightness features. We have applied the resulting new discriminator to an extensive data set of 0.3-meter resolution, HH polarization imagery gathered with the LL millimeter-wave SAR. The detection results are extremely good. In particluar, the new discriminator achieves a significant improvement in receiver operating characteristics, compared to an optimized version of the standard discriminator that is traditionally used in the LL ATR system. This result conclusively demonstrates that multiresolution methods have an effective and important role to play in SAR ATR algorithms.© (1995) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Abstract: Correlation filters have been successfully utilized for object detection in many applications. Each sensor type, however, presents different advantages and challenges. This paper describes the application of correlation filter techniques for automatic target recognition (ATR) to Laser Radar (LADAR) sensor images. Filters are designed using synthetic models, and incorporate range and aspect tolerance for mobile objects. The model generator easily takes into account the sensor field of view (FOV), look-down angle, ground cell size, and shadows. The filters are also designed to exploit various coordinate transforms that are feasible with a LADAR sensor. The correlation algorithm has the unique potential of exploiting the intensity information in conjunction with the range measurements provided by the LADAR. Examples using fixed and mobile targets are presented, along with statistical performance results.

Abstract: An approach to sea target recognition using an incoherent marine surveillance radar is examined in this paper. Reasons for reducing the target of recognition to its tonnage class, approximate size and resulting maneuvering abilities are described. An algorithm for automatic target recognition that effectively eliminates the negative impact of noise, sea clutter and distance from the target is described. Possibilities for application in sea navigation are proposed based on experimental findings.

Abstract: The most important property of a pattern recognition system is it's generalization capability. Previous research shows that neural networks generalize well and approximate arbitrarily complex functions. Feature extraction and decision-making are both important components. Unfortunately, designing effective feature extraction procedures is a very difficult task. For this reason, a heterogeneous neural network that can learn feature extraction and classification simultaneously is very attractive. The nonlinear correlation filter neural networks (NCFNN) learn correlation filters in the frequency domain and simultaneously learn a nonlinear combination of the outputs of the correlation filters for object detection. Two morphological neural networks, the generalized-mean morphology neural network (GMMNN) and the ordinary morphology neural network (OMNN), learn morphological structuring elements as feature extractors simultaneously with classification. They provide a general problem-independent methodology for designing morphological structuring elements. They perform feature extraction using a novel gray-scale Hit-Miss transform. The OMNN is invariant to shifts in gray-scale. The GMMNN and the OMNN were applied to pattern recognition problems. For binary handwritten digits, they produced performance comparable to that obtained using the ordinary linear shared-weight neural network (LSNN and NBLSNN) that has been used by others previously. However, the OMNN trained faster. For gray-scale patterns, the morphological neural networks produced superior performance. The LSNN, the OMNN and the NCFNN were applied to automatic target recognition (ATR) problems. Two data sets were used: Forward Looking Infrared image scenes of tanks and visual image scenes of a parking lot containing occluded vehicles. Several performance measurements and target-aim-point selection algorithms were defined. The OMNN performed significantly better than the other; especially at detecting occluded vehicles and reducing false alarm rates. All the networks performed significantly better than a Minimum-Average-Correlation-Energy filter technique. The new neural networks provide a general methodology for designing nonlinear filters without specific knowledge of the problem domain. They performed better than existing shared-weight network and matched filter approaches. In particular, the OMNN produced the best performance. It trained relatively quickly and is independent of shifts in gray-level. The NCFNN, the LSNN and the NBLSNN all produced similar performance rates. These networks are not problem-specific and can be widely used for other pattern recognition and ATR problems.

Abstract: Automatic target detection and recognition (ATR) requires the ability to optimally extract the essential features of an object from (usually) cluttered environments. In this regard, efficient data representation domains are required in which the important target features are both compactly and clearly represented, enhancing ATR. Since both detection and identification are important, multidimensional data representations and analysis techniques, such as the continuous wavelet transform (CWT), are highly desirable. First we review some relevant properties of two 2D CWT. Then we propose a two-step algorithm based on the 2D CWT and discuss its adequacy for solving the ATR problem. Finally we apply the algorithm to various images.

Abstract: Infrared imagery scenes change continuously with environmental conditions. Strategic targets embedded in them are often difficult to be identified with the naked eye. An IR sensor-based mine detector must include Automatic Target Recognition (ATR) to detect and extract land mines from IR scenes. In the course of the ATR development process, mine signature data were collected using a commercial 8-12 (mu) spectral range FLIR, model Inframetrics 445L, and a commercial 3-5 (mu) starting focal planar array FLIR, model Infracam. These sensors were customized to the required field-of-view for short range operation. These baseline data were then input into a specialized parallel processor on which the mine detection algorithm is developed and trained. The ATR is feature-based and consists of several subprocesses to progress from raw input IR imagery to a neural network classifier for final nomination of the targets. Initially, image enhancement is used to remove noise and sensor artifact. Three preprocessing techniques, namely model-based segmentation, multi-element prescreener, and geon detector are then applied to extract specific features of the targets and to reject all objects that do not resemble mines. Finally, to further reduce the false alarm rate, the extracted features are presented to the neural network classifier. Depending on the operational circumstances, one of three neural network techniques will be adopted; back propagation, supervised real-time learning, or unsupervised real-time learning. The Close Range IR Mine Detection System is an Army program currently being experimentally developed to be demonstrated in the Army's Advanced Technology Demonstration in FY95. The ATR resulting from this program will be integrated in the 21st Century Land Warrior program in which the mine avoidance capability is its primary interest.

Abstract: Pattern recognition involves the correct recognition of an object irrespective of rotation, scale and translation. In this paper the authors have come up with a recognition scheme, that has shown 100% recognition rate for all rotation, translation and tolerates a scale factor from 1/2 to 2. The use of the Fourier Mellin transform to get features invariant to rotation, scale and translation has been attempted previously. The contribution of this paper is in the use of neural networks to classify the invariant patterns obtained by the use of FMT, thereby providing robustness to the whole scheme. The efficiency of such a scheme can be judged by the high recognition rate obtained even for partially occluded images.

Abstract: The performance of automatic target recognition (ATR) systems using thermal infrared images is limited by the low contrast in intensity for terrestrial scenes. We are developing a thermal imaging technique where, in each image pixel, a combination of intensity and polarization data is captured simultaneously. In this paper, we demonstrate, using synthetic polarization images, that a combination of intensity and polarization data will significantly improve the performance of detection and classification functions in an ATR system. The images were generated using a ray tracing computer program, modified to calculate the polarization characteristics of thermal radiation emitted from surfaces. We developed novel polarization- sensitive target edge detection, segmentation, and recognition algorithms. A set of performance metrics for the evaluation showed that, for large ranges of viewing elevation and aspect angles, using a combination of polarization and intensity data significantly improves the performance of the algorithms over using only the intensity data.

Abstract: This paper presents a real time SAR/IR fusion system for automatic target recognition (ATR) applications. The system contains three major components: preprocessing, registration and fusion. The registration algorithm is based on Zheng-Chellappa's (see IEEE Transactions on Image Processing, no.7, p.311, 1993) APFBR work. The fusion algorithm employs a biologically inspired merging rule to carry out SAR/IR fusion at the pixel level. In the fused images, features from individual sensor images are not only preserved but also enhanced. The fused images are more suitable for visual perception and ATR operations. The system has been implemented on a SIMD parallel processor to achieve real-time performance. It has been extensively tested on real SAR/IR data.

Abstract: A potential target scenery image data is filtered and complex-multiplied h synthetic discriminant functions to produce a two-dimensional, cross-correlated surface. The surface is then analyzed to determine the spatial coordinates of the target, if any.

Abstract: Our pattern theoretic approach to the automated understanding of forward-looking infrared (FLIR) images brings the traditionally separate endeavors of detection, tracking, and recognition together into a unified jump-diffusion process New objects are detected and object types are recognized through discrete jump moves Between jumps, the location and orientation of objects are estimated via continuous diffusions An hypothesized scene, simulated from the emissive characteristics of the hypothesized scene elements, is compared with the collected data by a likelihood function based on sensor statistics This likelihood is combined with a prior distribution defined over the set of possible scenes to form a posterior distribution The jump-diffusion process empirically generates the posterior distribution Both the diffusion and jump operations involve the simulation of a scene produced by a hypothesized configuration Scene simulation is most effectively accomplished by pipelined rendering engines such as silicon graphics We demonstrate the execution of our algorithm on a silicon graphics onyx/reality engine

Abstract: The complementary nature of LADAR, FLIR and color data for Automatic Target Recognition (ATR) is being explored by new algorithms in a three stage recognition system. The stages are initial detection, target class and pose hypothesis generation, and precise model to multisensor coregistration matching. Coregistration globally aligns 3D target models with range, IR and color imagery while simultaneously refining registration parameters between sensors. This model directed approach is expected to improve ATR performance for occluded targets, targets seen at unusual angles, and targets in cluttered settings. Color is used for initial target detection under daylight conditions. Camouflage learned from training examples generalizes across vehicles and distinguishes targets from natural terrain. Target class and pose hypothesis generation will draw upon existing LADAR boundary matching work extended to tolerate more occlusion, clutter and viewpoint variation. New model to multisensor coregistration algorithms appear robust in early tests and are the basis for future coregistration matching. A new interactive 3D visualization environment allows inspection of multisensor data, coregistration, and monitoring of recognition.© (1995) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Abstract: We proposed an improved version of the SOFM/LVQ classifier currently used in an ATR system for SAR imagery. This classifier was originally designed to construct a few number of templates to represent a set of targets with different orientations. The classifier accepts an input of a target, computes distances of this data with those representative templates, and then classifies this data to the target class with the shortest distance. In this paper, we focus on the issue of how to identify and reject data from targets outside the given data set, such as man- made clutters. To reject clutters, we propose two discrimination functions, distance and entropy measures. With the distance discriminator, we have obtained a very good classification performance when all data are from the given target sets. However, the simple distance measure produces poor classification results when unknown targets such as natural or manmade clutters are present and when each target is represented by a small number of templates. We correct this deficiency by incorporating an entropy measure into the original classifier. With this entropy discriminator, our system rejects a majority of the false alarms while maintaining a high correct classification rate with a relatively few templates for each target. Although, this system was tested on real ISAR data and showed a very good performance, the data was obtained from `turntable' experiment with a fixed depression angle and known target location. One of the future research directions is to test this algorithm with real `field' SAR data and study the robustness of the system.

Abstract: The resolution capability of imaging systems is affected by blur resulting from vibration and motion during the exposure. This blur is often more severe than electronic and optical resolution limitations inherent in the system. Such image quality degradation must be considered when dealing with the development and analysis of automatic target recognition (ATR) systems. This research analyzes the influence of image vibrations and motion on the probability of acquiring a target with an ATR system. The analysis includes accepted metrics that characterize the relationship existing between the target and its background. A high level of correlation is expected between these factors and the probability of target detection permitting efficient performance in the prediction and evaluation of any ATR system. Such correlations are considered here in the presence of sensor motion and vibration and situations are considered in which the probability of recognition is improved by the motion, despite the blur. The results of this research can be implemented in military applications as well as in developing image restoration procedures for image-blur conditions.

Abstract: This paper presents an automatic target recognition (ATR) system for laser radar (LADAR) imagery, designed to classify objects at multiple levels of discrimination (target detection, classification, and recognition) from single LADAR images. Segmentation is performed in both the range and non-range LADAR channels and results combined to increase object detection rate or decrease false positive detection rate. Through use of the range data, object subimages are projected and rotated to canonical orientations, providing invariance to translation, scale and rotations in 3-D. Global features are extracted for rapid target detection and local receptive field features are computed for target recognition 100% detection and recognition rates are shown for a small set of real LADAR data.

Abstract: An automatic taget recognition (ATR) technique developed by the authors features analytically derived object models which are formed from entire image suites, yet are compact and allow a direct target recognition and pose determination procedure. In contrast to the pose-invariant information used to form the models in conventional approaches, view-dependent information is retained in the formation of the compact models for this new approach. All model-based ATR systems are confronted with the problem of image variation as a function of viewing angle. This problem can be addressed by use of an exhaustive library of views, at the expense of a large suite of literal images and a computationally intensive search-based recognition process. Means for overcoming these storage and processing obstacles have traditionally invloved some type of view-independent target representation, often developed from some composite view of the target over the viewing angles of interest. This results in a much more compact target model, and a more direct recognition process. Unfortunately, the gains in storage and computational requirements of these invariant algorithms come at the price of diminished target discrimination capability. The new algorithm incorporates pose as a fundamental parameter which is solved for as part of the recognition process, and does not discard the pose-related information which is relevant to target recognition. In this paper, the newly developed technique is applied to synthetic aperture radar images to develop receiver operating characteristic curves in the presence of both multiplicative noise and clutter. Comparative curves are also developed for a conventional generalized quandratic classifier ATR system.