This PDF file contains the front matter associated with SPIE Proceedings Volume 8136, including the Title Page, Copyright information, Table of Contents, Introduction, and Conference Committee listing.

Linear classifiers based on computation over the real numbers R (e.g., with operations of addition and multiplication) denoted by (R, +, x), have been represented extensively in the literature of pattern recognition. However, a different approach to pattern classification involves the use of addition, maximum, and minimum operations over the reals in the algebra (R, +, maximum, minimum) These pattern classifiers, based on lattice algebra, have been shown to exhibit superior information storage capacity, fast training and short convergence times, high pattern classification accuracy, and low computational cost. Such attributes are not always found, for example, in classical neural nets based on the linear inner product. In a special type of lattice associative memory (LAM), called a dendritic LAM or DLAM, it is possible to achieve noise-tolerant pattern classification by varying the design of noise or error acceptance bounds. This paper presents theory and algorithmic approaches for the computation of noise-tolerant lattice associative memories (LAMs) under a variety of input constraints. Of particular interest are the classification of nonergodic data in noise regimes with time-varying statistics. DLAMs, which are a specialization of LAMs derived from concepts of biological neural networks, have successfully been applied to pattern classification from hyperspectral remote sensing data, as well as spatial object recognition from digital imagery. The authors' recent research in the development of DLAMs is overviewed, with experimental results that show utility for a wide variety of pattern classification applications. Performance results are presented in terms of measured computational cost, noise tolerance, classification accuracy, and throughput for a variety of input data and noise levels.

In Bayesian pattern recognition research, static classifiers have featured prominently in the literature. A static classifier is essentially based on a static model of input statistics, thereby assuming input ergodicity that is not realistic in practice. Classical Bayesian approaches attempt to circumvent the limitations of static classifiers, which can include brittleness and narrow coverage, by training extensively on a data set that is assumed to cover more than the subtense of expected input. Such assumptions are not realistic for more complex pattern classification tasks, for example, object detection using pattern classification applied to the output of computer vision filters. In contrast, we have developed a two step process, that can render the majority of static classifiers adaptive, such that the tracking of input nonergodicities is supported. Firstly, we developed operations that dynamically insert (or resp. delete) training patterns into (resp. from) the classifier's pattern database, without requiring that the classifier's internal representation of its training database be completely recomputed. Secondly, we developed and applied a pattern replacement algorithm that uses the aforementioned pattern insertion/deletion operations. This algorithm is designed to optimize the pattern database for a given set of performance measures, thereby supporting closed-loop, performance-directed optimization. This paper presents theory and algorithmic approaches for the efficient computation of adaptive linear and nonlinear pattern recognition operators that use our pattern insertion/deletion technology - in particular, tabular nearest-neighbor encoding (TNE) and lattice associative memories (LAMs). Of particular interest is the classification of nonergodic datastreams that have noise corruption with time-varying statistics. The TNE and LAM based classifiers discussed herein have been successfully applied to the computation of object classification in hyperspectral remote sensing and target recognition applications. The authors' recent research in the development of adaptive TNE and adaptive LAMs is overviewed, with experimental results that show utility for a wide variety of pattern classification applications. Performance results are presented in terms of measured computational cost, noise tolerance, and classification accuracy.

Over the past quarter century, concepts and theory derived from neural networks (NNs) have featured prominently in the literature of pattern recognition. Implementationally, classical NNs based on the linear inner product can present performance challenges due to the use of multiplication operations. In contrast, NNs having nonlinear kernels based on Lattice Associative Memories (LAM) theory tend to concentrate primarily on addition and maximum/minimum operations. More generally, the emergence of LAM-based NNs, with their superior information storage capacity, fast convergence and training due to relatively lower computational cost, as well as noise-tolerant classification has extended the capabilities of neural networks far beyond the limited applications potential of classical NNs. This paper explores theory and algorithmic approaches for the efficient computation of LAM-based neural networks, in particular lattice neural nets and dendritic lattice associative memories. Of particular interest are massively parallel architectures such as multicore CPUs and graphics processing units (GPUs). Originally developed for video gaming applications, GPUs hold the promise of high computational throughput without compromising numerical accuracy. Unfortunately, currently-available GPU architectures tend to have idiosyncratic memory hierarchies that can produce unacceptably high data movement latencies for relatively simple operations, unless careful design of theory and algorithms is employed. Advantageously, some GPUs (e.g., the Nvidia Fermi GPU) are optimized for efficient streaming computation (e.g., concurrent multiply and add operations). As a result, the linear or nonlinear inner product structures of NNs are inherently suited to multicore GPU computational capabilities. In this paper, the authors' recent research in lattice associative memories and their implementation on multicores is overviewed, with results that show utility for a wide variety of pattern classification applications using classical NNs or lattice-based NNs. Dataflow diagrams are presented in terms of a parameterized model of data burden and LAM partitioning.

Digital watermarking continues to be an open area of research. In this work, fractals are employed to spatially embed the watermarks in the RGB domain. The watermarks are tested separately in each of the three planes R, G, and B. A blind detection scheme is utilized in which the only information required for detection is the fractals used for the embedding. Next, combinations of embedding in the RG, RB, GB, and RGB planes are used. The efficacy of the embedding combinations in the various planes are studied to determine the best combinations for the tested fractals, images and attacks. The results are compared with previously published methods by the authors.

We present a running approximation of discrete signals by a FIR filter bank that minimizes various worst-case measures of error, simultaneously. We assume that the discrete signal is a sampled version of unknown original band-limited signal that has a main lobe and small side-lobes. To restrict frequency characteristics of signals in this discussion, we impose two restrictions to a set of signals. One is a restriction to a weighted-energy of the Fourier transform of the discrete signal treated actually in the approximation and another is a restriction to a measure like Kullback-Leibler divergence between the initial analog signal and the final discrete approximation signal. Firstly, we show interpolation functions that have an extended band-width and satisfy condition called discrete orthogonality. Secondly, we present a set of signals and a running approximation satisfying all of these conditions of the optimum approximation.

The minimization of approximation errors in a FIR multi-dimensional filter bank for band-limited signals is the important problem of multi-dimensional signal processing. In this paper, we consider the interpolation approximation that is modeled as a certain multi-dimensional FIR filter bank. Firstly, we introduce the known optimum approximation of multi-dimensional signals in FIR filter banks using a finite number of sample values. Secondly, we explain briefly a new concept of multi-legged-type signal that is a combined-signal of many one-dimensional band-limited signals. Backbone of this multi-legged-type signal is constituted with a set of small separable segments of the above one-dimensional signals that are determined by the proposed running approximation. Thirdly, we extend this concept to multi-dimensional hyper signal space. Based on this concept, we present an approximation method of the extended multi-dimensional multi-legged-type signals and we prove that this approximation in the hyper domain is the optimum. Finally, we define measures of error that become the proposed measures of error in the position of the backbone made by the corresponding multi-dimensional running approximation and become small about the other errors. Based on these measures of error, we prove that the presented multi-dimensional optimum approximation minimizes various continuous worst-case measures of the running approximation error at the same time.

An overview of three-dimensional (3D) image recognition using photon-counting integral imaging is presented. A Poisson distribution model is used to generate photon-counting elemental images and reconstruct 3D images. Slices of 3D reconstructed images are the image source for recognition purpose. Experimental and computational results are shown to evaluate the performance of the method.

This paper introduces an algorithm for "mapping" locations in a 2D image of an object onto a 3D CAD model of that object. The goal is to align the image with the CAD model so that the location of various points in the image can be found on the object. We will actually present two methods for achieving this goal. One relies on projective invariants and requires no information about the camera parameters. Unfortunately, its performance is sometimes erratic in the face of particular types of errors. The other method assumes that the user is able to identify a minimum of 4 feature points in the image and the corresponding points on the CAD model. Beyond that, no information is assumed except for knowledge of the size of the camera's CCD and the location of the center of projection in the image. This second algorithm is highly accurate and robust in the face of real-world errors. We call this second algorithm the quaternion optimization algorithm. It positions the camera in CAD space so as to optimally align the matched features. This in turn permits the mapping of other image pixels to the CAD model.plica

In this paper, we define a multi-input multi-output system composed of given analysis-filter matrices, given sampler matrices and interpolation-filter matrices to be optimized, respectively. It is assumed that input-signal vectors of this system have a finite number of variables and these input-signal vectors are contained in a certain given set of input-signal vectors. Firstly, we define new notations which expresses a kind of product between two matrices or between a vector and a matrix. Using these new notations, we show that the presented approximation satisfies a certain two conditions and prove that the presented approximation minimizes any upper-limit measure of error compared to any other linear or nonlinear approximation with same sample values, simultaneously.

The minimization of the error associated with a running approximation by a filter bank is one of the most important problems of the signal processing. In this paper, for a set of vector-signals such that generalized Fourier transforms have weighted norms smaller than a given positive number, we present the extended optimum running approximation that minimizes various continuous worst-case measures of approximation error at the same time. In this discussion, we introduce a new concept of multi-legged-type signal that is a combined-signal of many one-dimensional band-limited signals. Backbone of this multi-legged-type signal is constituted with a series of small separable segments of the above one-dimensional signals that are determined by the proposed running approximation. Based on this concept, we propose an approximation method of the multi-legged-type signals and we prove that this approximation is the optimum. Then, we define measures of error that become the proposed measures of error in the position of the backbone made by the corresponding running approximation and become small about the other errors. Based on these measures of error, we prove that the presented extended optimum approximation minimizes various continuous worst-case measures of the running approximation error at the same time. As an application, multiple-input multiple-output/space division multiplexing system is discussed.

This paper proposes a novel recognition scheme algorithm for semantic labeling of 2D object present in still images. The principle consists of matching unknown 2D objects with categorized 3D models in order to infer the semantics of the 3D object to the image. We tested our new recognition framework by using the MPEG-7 and Princeton 3D model databases in order to label unknown images randomly selected from the web. Results obtained show promising performances, with recognition rate up to 84%, which opens interesting perspectives in terms of semantic metadata extraction from still images/videos.

A hyperspectral imaging system has been set up and used to capture hyperspectral image cubes from various samples in the 400-1000 nm spectral region. The system consists of an imaging spectrometer attached to a CCD camera with fiber optic light source as the illuminator. The significance of this system lies in its capability to capture 3D spectral and spatial data that can then be analyzed to extract information about the underlying samples, monitor the variations in their response to perturbation or changing environmental conditions, and compare optical properties. In this paper preliminary results are presented that analyze the 3D spatial and spectral data in reflection mode to extract features to differentiate among different classes of interest using biological and metallic samples. Studied biological samples possess homogenous as well as non-homogenous properties. Metals are analyzed for their response to different surface treatments, including polishing. Similarities and differences in the feature extraction process and results are presented. The mathematical approach taken is discussed. The hyperspectral imaging system offers a unique imaging modality that captures both spatial and spectral information that can then be correlated for future sample predictions.

This paper presents a novel method for retrieving different instances of a region or object of interest in a video database. We propose three dynamic region construction and matching algorithms aiming at obtaining the most similar instance of the query model from each candidate image. The first two involve a greedy, dynamic region construction method. The third is based on simulated annealing, and aims at determining a global optimum. An interactive selection mechanism allows the user to select an object directly from the video and to start a query using as input this information in order to access visually similar content. Experimental results show promising performances, with object detection rates of up to 81%.

When performing image analysis, one of the most critical steps is the selection of appropriate techniques. A huge amount of features can be extracted from several techniques and the selection is commonly performed based on expert knowledge. In this paper we present the theory of experimental designs as a tool for an objective selection of techniques in image analysis domain. We present a study case for evaluating appearance retention in textile floor coverings using texture features. The use of experimental design theory permitted to select an optimal set of techniques for describing the texture changes due to degradation.

Edge detection process plays an important role in image processing, and at its most basic level classifies image pixels into edges and non-edge pixels. The accuracy of edge detection methods in general image processing determines the eventual success or failure of computerized analysis procedures which follow the initial edge detection determinations. In view of this downstream impact on pattern processing, considerable care should be taken to improve the accuracy of the front-end edge detection. In general, edges would be considered as abrupt changes or discontinuity in intensity of an image. Therefore, most of edge detection algorithms are designed to capture signal discontinuities but the spatial character of especially complex edge patterns has not received enough attention. Edges can be divided into basic patterns such as ramp, impulse, and step: different types have different shapes and consequent mathematical properties. In this paper, the behavior of various edge patterns, under different order derivatives in the discrete domain, are examined and analyzed to determine how to accurately detect and localize these edge patterns, especially reducing double edge response that is one important drawback to the derivative method. General rules about the depiction of edge patterns are proposed. Asides from the ideal patterns already described, other pattern types, such as stair and roof, are examined to broaden the initial analysis. Experiments conducted to test my propositions support the idea that edge patterns are instructive in enhancing the accuracy of edge detection and localization.

In this work we conduct a comparative study on different data compression methods applied to high resolution images of the solar surface acquired at the Solar Dunn Telescope in Sacramento Peak with the IBIS (Interferometric Bidimensional Spectrometer) instrument. Our aim is to perform an estimation of the quality, efficiency and workload of the considered computing techniques both in the so-called lossless modality, where in the reconstruction phase there is no loss of information, and in lossy mode, where it should be possible to reach a high compression ratio at the expense of image information. In the latter case we quantify the quality with image analysis conventional methods and more specifically with the reconstruction of physical parameters through standard procedures used in this kind of observations. The considered methods constitute the most frequently adopted image compression procedures in a variety of fields of application; they exploit in different ways the properties of the Discrete Wavelet Transforms often coupled with standard entropy coders or similar coding procedures applied to the different bit planes in order to allow a progressive handling of the original image. In the lossless approach we found that all methods give a compression ratio around 2. For a lossy compression we reached a compression ratio of 8 (equivalent to a 2 bit per pixel) without any perceptual difference between original and reconstructed images, but with effects on the photometric accuracy. We also tested the performance of 3-D lossy methods for the compression of data-cubes. Maintaining the same data degradation level, those methods allows us to gain a 2x in the compression ratio over the 2-D methods.

Distributed Fiber Vibrant Sensor System is a new type of system, which could be used in long-distance, strong-EMI condition for monitoring vibration and sound signals. Position determination analysis toward this system is popular in previous papers, but pattern recognition of the output signals of the sensor has been missed for a long time. This function turns to critical especially when it is used for real security project in which quick response to intrusion is a must. After pre-processing the output signal of the system, a MFCC-based approach is provided in this paper to extract features of the sensing signals, which could be used for pattern recognition in real project, and the approach is proved by large practical experiments and projects.

Computer aided surgery is by sure a set of technologies that provide a real support to surgeons during their operational job. This includes - but it is not limited to - novel sensors and systems and software for data analysis and visualization. In particular the use of intraoperational probes is eased if the position of the probe within the operational field can be exactly calculated by the supporting software. Commercial systems have already been developed for this purpose but their complexity and cost reduces their usability for the majority of probes. This paper presents a simple approach to calculate the probe position within the operational field that requires a minimum cost.