One typical problem in image processing is to remove noise. This problem becomes rather complicated if not one type of noise is present, i.e. if, for example, an image is corrupted by mixed noise. Non-adaptive nonlinear filters often do not provide desirable image processing quality. Nonlinear locally adaptive hard-switching filters are more flexible and effi-cient in this sense. However, for mixed noise case, all component filters employed in locally adaptive filtering (LAF) framework have to possess robust properties. In particular, this relates to the so-called noise suppressing filter (NSF) to be applied in image homogeneous regions. Below we show that myriad filter with properly selected tunable (lineariza-tion) parameter is able to be an efficient NSF that outperforms well known α-trimmed and Wilcoxon filters. For this purpose, general statistical analysis for myriad estimate of distribution shift (location) parameter is performed first. Then numerical simulation results for myriad estimator are obtained and compared to other robust shift parameter estimators for different parameters of data sample model used. The recommendations for tunable parameter selection are given. The proposed filtering technique is verified for both artificial and real images.

We present a multiscale region-merging segmentation algorithm based on nonlinear diffusion equations. The algorithm is applicable to vector-valued images such as color images, or feature images obtained by pre-processing a texture image. The algorithm is experimentally shown to provide accurate segmentations for texture images.

During critical situations, the precise digital processing of medical signals such as heartbeats is essential. Outside noise introduced into this data can lead to misinterpretation. Thus, it is important to be able to detect and correct the signal quickly and efficiently using digital filtering algorithms. With filtering, the goal is to remove noise locations by correctly identify the corrupted data points and replacing these locations with acceptable estimations of the original values. However, one has to be careful throughout the filtering process not to also eliminate other important detailed information from the original signal. If the filtered output is to be analyzed post-filtering, say for feature recognition, it is important that both the structure and details of the original clean signal remain. This paper presents an original algorithm and two variations, all using the logical transform, that strive to do this accurately and with low levels of computation. Using real heartbeat signals as test sets, the output is compared to that produced by median type filters, and results demonstrated over a variety of noise levels.

Reliable and accurate methods for road network detection and classification in satellite imagery are essential to many applications. We present an image vectorization approach to the road network extraction from digital imagery that is based on proximity graph analysis. An input to the presented approach is spectrally segmented image that contains a set of candidate road fragments. First, non-intersecting contours are extracted around image elements. Second, constrained Delaunay triangulation and Chordal Axis transform are used to extract global centerline topology characterization of the delineated candidate road fragments. Then, constrained Delaunay triangulation of the extracted set of attributed center lines is performed. The tessellation grid of the Delaunay triangulation covers the set of candidate road fragments and is adapted to its structure, since triangle vertices and edges reflect the shapes and spatial adjacency of the segmented regions. The produced Delaunay network edges can be attributed with spectral and structural characteristics that are used for spatial analysis of the edges relations. This leads to the reconstruction of the road network out of the Delaunay edges. A subset of the tessellation grid contains the Euclidian Minimum Spanning Tree that provides an approximation of road network. The approach can be generalized to the multi-criteria MST and multi-criteria shortest path algorithms to integrate other factors important for road network extraction, in addition to proximity relations considered by standard MST.

This paper provides an overview of the state-of-the-art techniques recently developed within the emerging field of dynamic mesh compression. Static encoders, wavelet-based schemes, PCA-based approaches, differential temporal and spatio-temporal predictive techniques, and clustering-based representations are considered, presented, analyzed, and objectively compared in terms of compression efficiency, algorithmic and computational aspects and offered functionalities (such as progressive transmission, scalable rendering, computational and algorithmic aspects, field of applicability...). The proposed comparative study reveals that: (1) clustering-based approaches offer the best compromise between compression performances and computational complexity; (2) PCA-based representations are highly efficient on long animated sequences (i.e. with number of mesh vertices much smaller than the number of frames) at the price of prohibitive computational complexity of the encoding process; (3) Spatio-temporal Dynapack predictors provides simple yet effective predictive schemes that outperforms simple predictors such as those considered within the interpolator compression node adopted by the MPEG-4 within the AFX standard; (4) Wavelet-based approaches, which provide the best compression performances for static meshes show here again good results, with the additional advantage of a fully progressive representation, but suffer from an applicability limited to large meshes with at least several thousands of vertices per connected component.

This paper investigates how the 3D facial animation techniques can be exploited within the specific framework of 2D cartoon production. An overview of the most representative 3D facial animation techniques is first presented. Physical modeling, free form deformations, direct face parameterizations and controller-based approaches are identified and discussed in detail. From this critical analysis of the literature, we selected for evaluation purposes two different controller-based approaches: RBF- and wire-based deformations. Experimental results have been carried out on a corpus of 3D face models with both neutral and target expressions available. The RBF-based techniques provide smoother and more stable deformation fields at a lower modeling effort than the wire-based approaches. Both methods are appropriate for achieving 2D/3D deformation and automating the 2D cartoon production.

Kolmogorov's structure function (KSF) is used in the algorithmic theory of complexity for describing the structure of a string by use of models (programs) of increasing complexity. Recently, inspired by the structure function, an extension of the minimum description length theory was introduced for achieving a decomposition of the total description of the data into a noise part and a model part, where the models are parametric distributions instead of programs, the code length necessary for the model part being restricted by a parameter. In this way a new "rate-distortion" type of curve is obtained, which may be further used as a general model of the data, quantifying the amount of noise left "unexplained" by models of increasing complexity. In this paper we present a complexity-noise function for a class of hierarchical image models in the wavelet transform domain, in the spirit of the Kolmogorov structure function. The minimization of the model description can be shown to have a form similar to one resulting from the minimization in the rate-distortion sense, and thus it will be achieved as in lossy image compression. As an application of the complexity-noise function introduced we study the image denoising problem and analyze the conditions under which the best reconstruction along the complexity-noise function is obtained.

Nowadays, alongside with the traditional voice signal, music, video, and 3D characters tend to become common data to be run, stored and/or processed on mobile phones. Hence, to protect their related intellectual property rights also becomes a crucial issue. The video sequences involved in such applications are generally coded at very low bit rates. The present paper starts by presenting an accurate statistical investigation on such a video as well as on a very dangerous attack (the StirMark attack). The obtained results are turned into practice when adapting a spread spectrum watermarking method to such applications. The informed watermarking approach was also considered: an outstanding method belonging to this paradigm has been adapted and re evaluated under the low rate video constraint. The experimental results were conducted in collaboration with the SFR mobile services provider in France. They also allow a comparison between the spread spectrum and informed embedding techniques.

Adaptive steganography is a statistical approach for hiding the digital information into another form of digital media. The goal is to ensure the changes introduced into the cover image remain consistent with the natural noise model associated with digital images. There are generally two classes of steganography − global and local. The global class encompasses all non-adaptive techniques and is the simplest to apply and easiest to detect. The second classification is the local class, which defines most of the present adaptive techniques. We propose a new adaptive technique that is able to overcome embedding capacity limitations and reduce the revealing artifacts that are customarily introduced when applying other embedding methods. To obtain the objectives, we introduce a third faction which is the pixel focused class of steganography. Applying a new adaptive T-order statistical local characterization, the proposed algorithm is able to adaptively select the number of bits to embed per pixel. Additionally, a histogram retention process, an evaluation measure based on the cover image and statistical analysis allow for the embedding of information in a manner which ensures soundness from multiple statistical aspects. Based on the results of simulated experiments, our method is shown to securely allow an increased amount of embedding capacity, simultaneously avoiding detection by varying steganalysis techniques.

In this paper we examine the problem of single-view recognition for sets of line features under generalized weak perspective projection. Our ultimate goal is to understand metrically the shape spaces for various labeled and unlabeled feature sets of lines and to represent those shapes using global shape coordinates. The later are "invariants" that provide an isometric embedding of the shape space in question (be it the object space or the image space) into an ambient Euclidean or projective space. The metrics involved are natural variants of the Procrustes metric of classical shape theory that is used in object recognition and image understanding tasks. In this first paper of what will be a series, we derive a fundamental set of equations that express the relationship between the 3D geometry of our lines and its "residual" in a 2D generalized weak perspective image. These equations are known as object/image relations. They completely and invariantly describe the mutual 3D/2D constraints, and can be exploited in a number of ways. For example, from a given 2D configuration of edges (lines) in an image, we can determine a set of nonlinear constraints on the geometric invariants of all 3D configurations capable of producing the given 2D configuration, and thus arrive at a test for determining the object being viewed. Conversely, given a 3D geometric configuration of lines (features on an object), we can derive a set of equations that constrain the invariants of the images of that object; helping us to determine if that particular object appears in various images.

In the framework of preoperative evaluation of the hepatic venous anatomy in living-donor liver transplantation or oncologic rejections, this paper proposes an automated approach for the 3D segmentation of the liver vascular structure from 3D CT hepatic venography data. The developed segmentation approach takes into account the specificities of anatomical structures in terms of spatial location, connectivity and morphometric properties. It implements basic and advanced morphological operators (closing, geodesic dilation, gray-level reconstruction, sup-constrained connection cost) in mono- and multi-resolution filtering schemes in order to achieve an automated 3D reconstruction of the opacified hepatic vessels. A thorough investigation of the venous anatomy including morphometric parameter estimation is then possible via computer-vision 3D rendering, interaction and navigation capabilities.

Quantitatively assessing myocardial perfusion is a key issue for the diagnosis, therapeutic planning and patient follow-up of cardio-vascular diseases. To this end, perfusion MRI (p-MRI) has emerged as a valuable clinical investigation tool thanks to its ability of dynamically imaging the first pass of a contrast bolus in the framework of stress/rest exams. However, reliable techniques for automatically computing regional first pass curves from 2D short-axis cardiac p-MRI sequences remain to be elaborated. We address this problem and develop an unsupervised four-step approach comprising: (i) a coarse spatio-temporal segmentation step, allowing to automatically detect a region of interest for the heart over the whole sequence, and to select a reference frame with maximal myocardium contrast; (ii) a model-based variational segmentation step of the reference frame, yielding a bi-ventricular partition of the heart into left ventricle, right ventricle and myocardium components; (iii) a respiratory/cardiac motion artifacts compensation step using a novel region-driven intensity-based non rigid registration technique, allowing to elastically propagate the reference bi-ventricular segmentation over the whole sequence; (iv) a measurement step, delivering first-pass curves over each region of a segmental model of the myocardium. The performance of this approach is assessed over a database of 15 normal and pathological subjects, and compared with perfusion measurements delivered by a MRI manufacturer software package based on manual delineations by a medical expert.

Hiding messages in image data, called steganography, is used for both legal and illicit purposes. The detection of hidden messages in image data stored on websites and computers, called steganalysis, is of prime importance to cyber forensics personnel. Automating the detection of hidden messages is a requirement, since the shear amount of image data stored on computers or websites makes it impossible for a person to investigate each image separately. This paper describes research on a prototype software system that automatically classifies an image as having hidden information or not, using a sophisticated artificial neural network (ANN) system. An ANN software package, the ISU ACL NetWorks Toolkit, is trained on a selection of image features that distinguish between stego and nonstego images. The novelty of this ANN is that it is a blind classifier that gives more accurate results than previous systems. It can detect messages hidden using a variety of different types of embedding algorithms. A Graphical User Interface (GUI) combines the ANN, feature selection, and embedding algorithms into a prototype software package that is not currently available to the cyber forensics community.

Intensity-based Non Rigid Registration (NRR) techniques using statistical similarity measures have been widely used to address mono- and multimodal image alignment problems in a robust and segmentation-free way. In these approaches, registration is achieved by minimizing the discrepancy between luminance distributions. Classical similarity criteria, including mutual information, f-information and correlation ratio, rely on global luminance statistics over the whole image domain and do not incorporate spatial information. This may lead to inaccurate or geometrically inconsistent (though visually satisfying) alignment of homologous image structures, making these criteria unreliable for atlas-based segmentation purposes. This paper addresses these limitations and presents a region-driven approach to statistical NRR based on regional non-parametric estimates of luminance distributions. The latter are derived from a regional segmentation of the target image which is used as a fixed object/scene template and induces regionalized statistical similarity measures. We provide the expressions of these criteria in the case of generalized information measures and correlation ratio, and derive the corresponding gradient flows over parametric and non-parametric transforms spaces. This approach is then applied to the joint non rigid segmentation and registration of short-axis cardiac perfusion MR sequences using a bi-ventricular heart template. In this framework, region-driven NRR allows for compensating for respiratory/cardiac motion artifacts, and fitting a segmental heart model used for quantitatively assessing regional myocardial perfusion. Experiments have been performed on a 15 pathological subjects database, demonstrating the relevance of region-driven NRR over global NRR in terms of computational performance and registration accuracy with respect to an expert reference.

Atomic decompositions are lower-cost alternatives to the principal component analysis (PCA) in tasks where sparse signal representation is required. In pattern classifications tasks, e.g. face detection, a careful selection of atoms is needed in order to ensure an optimal and fast-operating decomposition to be used in the feature extraction stage. In this contribution, adaptive boosting is used as criterion for selecting optimal atoms as features in frontal face detection system. The goal is to speed up the learning process by a proper combination of a dictionary of atoms and a weak learner. Dictionaries of anisotropic wavelet packets are used where the total number of atoms is still feasible for large-size images. In the adaptive boosting algorithm a Bayesian classifier is used as a weak learner instead of a simple threshold, thus ensuring a higher accuracy for slightly increased computational cost during the detection stage. The experimental results obtained for four different dictionaries are quite promising based on the good localization properties of the anisotropic wavelet packet functions.

In this paper, we compare two correlation techniques for face recognition. An optical correlator using the Binary Phase-Only Filter (BPOF) was simulated and tested on a database of faces. Its performance is compared with that of an extended correlation method suggested by Kyatkin and Chirikjian. The results are compared in terms of the number of false positives. For our tests and for both methods we use the challenging AT&T database of images with varying facial illuminations, facial expressions, and pose. We show how, from the accuracy point of view, the extended correlation method outperforms the BPOF.

The aim of this paper that is organized in three parts is to introduce the concept of parallel access of data in regular but not orthogonal grids. Although the orthogonal grid and the corresponding sampling methods are well-known for many years and well established in science and technology, there is a certain interest in 2- and 3-dimensional imaging to study trigonal and hexagonal grids. In the 2-dimensional case these grids are generated by tesellation of the plane using triangles and hexagons, respectively. They form very regular patterns and they have very nice properties according to the number of neighborhood pixels and distance values in electronic imaging. Moreover, it is known for a long time that the retina part of the human visual system can be modeled by a hexagonal packing structure of rods and cones. In this paper we study the connection and the influence of the necessary data structures, access patterns, and system architecture to model imaging algorithms with trigonal and hexagonal grids. In particular, we study the parallel access to straight lines and hexagonal "circles". We show a possible parallel memory architecture for the parallel conflict-free access to rows, straight lines and hexagonal "circles". The necessary fundamental notions are given in this part.

Industrial product quality is frequently assessed using up to second-order statistics of populations of measurements. Lately, a fuzzy interval number (FIN) was used for representing a whole population of samples. It turns out that a FIN can asymptotically capture statistics of all orders. The space F of FINs, including both conventional (fuzzy) numbers and conventional intervals, is studied here. A FIN is interpreted as a (linguistic) information granule that can capture industrial ambiguity. Based on generalized interval analysis it is shown rigorously that F is a metric mathematical lattice; moreover it is shown that F a cone in a linear space. An enhanced extension of Kohonen's Self-Organizing Map (KSOM), namely granular SOM or grSOM for short, is presented in F^N for inducing a distribution of FINs from populations of measurements. The grSOM produces descriptive decision-making knowledge (i.e. rules) from the training data by expert attaching labels to induced n-tuples of FINs. Generalization is feasible beyond rule support. A positive valuation function, computable genetically, can introduce tunable nonlinearities. Preliminary results are demonstrated regarding industrial fertilizer quality assessment. Fuzzy-mathematical-morphology-based image processing techniques, which combine binary thresholding and object recognition, are used to automatically measure the geometry of fertilizer granules. Additional measurements are also considered. The far-reaching practical potential of the proposed techniques is discussed.

In the present paper, lattice dendrite computation is extended with non-linear transformations of neural inputs that are applied before local discrimination is performed by each dendrite of an artificial neuron. At the expense of increasing the gap with biological analogies or biophysical similarities, the proposed mathematical extension to the basic single layer lattice perceptron model has the advantage that with appropriate input transformations one type synaptic connections can be used, excitatory or inhibitory only; similarly, a reduction in the number of dendrites needed to solve certain one-class recognition problems can be achieved. Illustrative examples are given to show the new capabilities and possible applications of this enhanced single layer lattice perceptron.

Content based image retrieval (CBIR) systems are database management systems that emply features extracted from the image as the indices used in the search of the database. Images are retrieved on the basis of the similarity with the query image. Indexing hyperspectral images is a special case of CBIR, with the added complexity of the high dimensionality of the pixels. We propose the use of endmembers as the hyperspectral image characterization. We thus define a similarity measure between hyperspectral images based on these image endmembers. The endmembers must be induced from the image data in order to automate the process. Enmembers can be assumed to be morphologically independent, a notion originally introduced to study the noise robustnes of Morphological Networks. For this induction we use Associative Morphological Memories (AMM) as detectors of Morphological Independence conditions.

Morphological associative memories (MAM's) belong to a class of artificial neural networks that perform the operations erosion or dilation of mathematical morphology at each node. Therefore we speak of morphological neural networks. Alternatively, the total input effect on a morphological neuron can be expressed in terms of lattice induced matrix operations in the mathematical theory of minimax algebra. Neural models of associative memories are usually concerned with the storage and the retrieval of binary or bipolar patterns. Thus far, the emphasis in research on morphological associative memory systems has been on binary models, although a number of notable features of autoassociative morphological memories (AMM's) such as optimal absolute storage capacity and one-step convergence have been shown to hold in the general, gray-scale setting. In previous papers, we gained valuable insight into the storage and recall phases of AMM's by analyzing their fixed points and basins of attraction. We have shown in particular that the fixed points of binary AMM's correspond to the lattice polynomials in the original patterns. This paper extends these results in the following ways. In the first place, we provide an exact characterization of the fixed points of gray-scale AMM's in terms of combinations of the original patterns. Secondly, we present an exact expression for the fixed point attractor that represents the output of either a binary or a gray-scale AMM upon presentation of a certain input. The results of this paper are confirmed in several experiments using binary patterns and gray-scale images.

This article introduces a new theoretical framework to describe the behavior of the Steinbuch's Lernmatrix. The properties of this old associative memory can be modeled using set theory and order relationships, analogously to morphological associative memories. The obtained results allow the Lernmatrix, four decades before its creation, to be a good alternative for pattern classification and recognition.

In this work, we are proposing a model that improves closed loop automatic controllers for real time processes, using only a positive function to deal with the process variations, no matter if they are positive or negative, the unit uses the sign to deal with the control output value. The first part of this document, gives an introduction to the kind of control systems that can be applied, and the reason to make this project. The second part gives the parameters to be used on the control model, as well as the signal conditioning stage. Then at third part, the design of the model is explained, followed by Results and conclusions at fourth and fifth parts respectively.

This proposal presents a novel use of Weightless Neural Networks (WNN) and Steinbuch Lernmatrix for pattern recognition and classification. High speed of learning, easy of implementation and flexibility given by WNN, combined with the learning capacity, recovery efficiency, noise immunity and fast processing shown by Steinbuch Lernmatrix are key factors considered on the pattern recognition exposed by the suggested model. For experimental purposes, the fundamental pattern sets are built and provided to the model under the learning phase. The additive, subtractive and mixed noises are applied to fundamental patterns to check out the response of the model during the recovery phase. Field Programmable Gate arrays are used in the implementation of such model, since it allows custom user-defined models to be embedded in a reconfigurable hardware platform, and provides block memories and dedicated multipliers suitable for the model.

This paper presents a novel two-layer feedforward neural network that acts as an associative memory for pattern recall. The neurons of this network have dendritic structures and the computations performed by the network are based on lattice algebra. Use of lattice computation avoids multiplicative processes and, thus, provides for fast computation. The synaptic weights of the axonal fibers are preset, making lengthy training unnecessary. The proposed model exhibits perfect recall for perfect input vectors and is extremely robust in the presence of noisy or corrupted input.

Among various factors degrading captured image quality, motion blur is the most common. The user becomes aware of the blur only after viewing pictures on a higher resolution display. The causes of motion blur are object movement, camera shake, or any relative speed between the object and the camera. To avoid this problem, many anti-shaking or image stabilization techniques have been developed. However, a detecting mechanism for motion blur is still lacking. Hence, this paper will address some possible solutions and evaluate their performance. The purpose of a motion blur detector is to classify the digital image as blurred or clear and inform users. This function can supply information for users to decide to retake the picture immediately instead of turning the camera to playback mode to check. For achieving higher error tolerance and adaptation to different image capturing circumstances, a machine learning technique is employed. Different digital image processing schemes are explored to find the most discriminative features. Among the many machine techniques, Support Vector Machine (SVM) has been implemented. To achieve the best performance for SVM, inherent information extraction from motion blurred images is extremely important. Thus, several signal transformations including discrete Fourier, discrete cosine, and Radon transformation have been explored. A comparison of the performance of different feature vectors, kernel function, and parameters will also been addressed in this paper.

We develop a new methodology for constructing hierarchical stochastic image models called spatial random trees (SRTs) which admit polynomial-complexity exact inference algorithms. We use our framework of multitree dictionaries as the starting point for this construction. We develop an efficient algorithm for computing the EM updates and use it to estimate the model parameters. We illustrate our models and algorithms through image classification experiments.

Pattern classification theory involves an error criterion, optimal classifiers, and a theory of learning. For clustering, there has historically been little theory; in particular, there has generally (but not always) been no learning. The key point is that clustering has not been grounded on a probabilistic theory. Recently, a clustering theory has been developed in the context of random sets. This paper discusses learning within that context, in particular, k- nearest-neighbor learning of clustering algorithms.

The MicroMak device is a new, high-precision, very compact star sensor weighing less than 100 grams, with three independent 4-degree square fields of view. The collection telescope is a Maksutov design that incorporates three telescopes into a single sensor head. The sensor is designed for star identification and spacecraft attitude determination with a device that offers unprecedented low cost, volume and mass. While star trackers have achieved sub-arcsecond accuracy by utilizing sophisticated algorithms and complex hardware, the MicroMak sensor must rely on fairly efficient algorithms that utilize data from only the image sensor. This paper will discuss the attitude determination algorithm as well as a complete end-to-end simulation of the system that was used to optimize the design and predict performance. This simulation accepts various star and sensor parameters as inputs, and generates error estimates of attitude of the sensor. The inputs include color temperatures and magnitudes of stars, focal length, receiver aperture, reflectivity curves of mirrors, modulation transfer function of the telescope system, vignetting effects, jittter characteristics, spacecraft spin rate and spin axis, detector pixel size, read noise, dark noise, sensor update rate, quantum efficiency as a function of wavelength, and detector fill factor. A complete forward model of the optical train has been built, and used with a maximum likelihood estimator to generate estimates of sensor attitude. A Monte Carlo algorithm was used to generate error distributions on the attitude error given the noise and distortions injected into the measurement.

Error estimation is a key aspect of statistical pattern recognition. The true classification error rate is usually unavailable since it depends on the unknown feature-label distribution. Hence, one needs to estimate the error rate from the available sample data. This paper presents a concise, mathematically rigorous review of the subject of error estimation in statistical pattern recognition, pointing to the pitfalls that arise in small-sample settings due to the use of "rules of thumb" and a neglect for proper mathematical understanding of the problem.

Object boundary extraction from binary images is important for many applications, e.g., image vectorization, automatic interpretation of images containing segmentation results, printed and handwritten documents and drawings, maps, and AutoCAD drawings. Efficient and reliable contour extraction is also important for pattern recognition due to its impact on shape-based object characterization and recognition. The presented contour tracing and component labeling algorithm produces dilated (sub-pixel) contours associated with corresponding regions. The algorithm has the following features: (1) it always produces non-intersecting, non-degenerate contours, including the case of one-pixel wide objects; (2) it associates the outer and inner (i.e., around hole) contours with the corresponding regions during the process of contour tracing in a single pass over the image; (3) it maintains desired connectivity of object regions as specified by 8-neighbor or 4-neighbor connectivity of adjacent pixels; (4) it avoids degenerate regions in both background and foreground; (5) it allows an easy augmentation that will provide information about the containment relations among regions; (6) it has a time complexity that is dominantly linear in the number of contour points. This early component labeling (contour-region association) enables subsequent efficient object-based processing of the image information.

The assessment of bronchial reactivity and wall remodeling in asthma plays a crucial role in better understanding such a disease and evaluating therapeutic responses. Today, multi-detector computed tomography (MDCT) makes it possible to perform an accurate estimation of bronchial parameters (lumen and wall areas) by allowing a quantitative analysis in a cross-section plane orthogonal to the bronchus axis. This paper provides the tools for such an analysis by developing a 3D investigation method which relies on 3D reconstruction of bronchial lumen and central axis computation. Cross-section images at bronchial locations interactively selected along the central axis are generated at appropriate spatial resolution. An automated approach is then developed for accurately segmenting the inner and outer bronchi contours on the cross-section images. It combines mathematical morphology operators, such as "connection cost", and energy-controlled propagation in order to overcome the difficulties raised by vessel adjacencies and wall irregularities. The segmentation accuracy was validated with respect to a 3D mathematically-modeled phantom of a pair bronchus-vessel which mimics the characteristics of real data in terms of gray-level distribution, caliber and orientation. When applying the developed quantification approach to such a model with calibers ranging from 3 to 10 mm diameter, the lumen area relative errors varied from 3.7% to 0.15%, while the bronchus area was estimated with a relative error less than 5.1%.

For a PET scanner with circular array of detectors, the width of line-of-response (LOR) decreases as the distance between the LOR and the center increases. The decrease of width of the LOR leads to problem of non-uniform and under sampling of projections. The consequence of non-uniform sampling is the distortion of high frequency reconstructed images or loss of fine detail. Correcting this non-uniform sampling problem is known as arc-correction. The purpose of this study is to create the best estimate of non-uniformly sampled projections from uniformly spaced sets of LOR. Four polynomial type interpolating algorithms: Lagrange, iterative Neville, natural cubic spline and clamped cubic spline are used to get the best estimate of projections. A set of simulated projections are generated. The simulated projections are divided into two sets: the first set has 10 functions of pulses such that f¹¹ has one pulse, f¹² has two pulses and so on. In the second set f²¹ has one triangular pulse, f²² has two triangular pulses and so on. For each group interpolated data is compared to the original data. In addition, one projection of a 20cm FDG filled disk is used for comparison with simulated data. It is shown that clamped and natural cubic spline accuracy is superior to the other three algorithms in every case but Lagrange outperforms other algorithms for the speed of execution.

A solution for the problem of nonsupervized recognition in the conditions of a priori indefinite number of object classes in radar images is presented. The designed algorithm performs image clustering to divide image objects into classes. The region of interest is can be chosen by user and then probabilistic filtering is applied to recognize the objects of the predetermined class on the entire image. The algorithm is operated on the multichannel data and shows stable recognition results.