Transform methods have played an important role in signal and image processing applications. Recently, Selesnick has constructed the new orthogonal discrete wavelet transform, called the slantlet wavelet, with two zero moments and with improved time localization. The discrete slantlet wavelet transform is carried out by an existing filterbank which lacks a tree structure and has a complexity problem. The slantlet wavelet has been successfully applied in compression and denoising. In this paper, we present a new class of orthogonal parametric fast Haar slantlet transform system where the slantlet wavelet and Haar transforms are special cases of it. We propose designing the slantlet wavelet transform using Haar slantlet transform matrix. A new class of parametric filterbanks is developed. The behavior of the parametric Haar slantlet transforms in signal and image denoising is presented. We show that the new technique performs better than the slantlet wavelet transform in denoising for piecewise constant signals. We also show that the parametric Haar slantlet transform performs better than the cosine and Fourier transforms for grey level images.

In this paper a general parameterized integer-to-integer discrete cosine transform (DCT) is introduced. The parameter of the transform relates to the operation of floor function. The traditional method of integer transforms is based on the use of floor function by adding number 0.5 in each stage of the integer transform calculation. We consider the integer DCT with other parameters to be chosen in an optimal way. The optimality is with respect to the minimal mean-square-error of the integer DCT compared with the original DCT with the float-point multiplication. And while achieving the minimum error we are able to reconstruct the inputs exactly. Examples for the 2-, 4-, and 8-point integer reversible and inverse DCTs of types II and IV are analyzed in detail and optimal parameters for estimation of these transforms by integer DCTs are defined.

The presented paper introduces a new class of wavelets that includes the simplest Haar wavelet (Daubechies-2) as well as the Daubechies-4 wavelet. This class is shown to have several properties similar to the Daubechies wavelets. In application, the new class of wavelets has been shown to effectively denoise ECG signals. In addition, the paper introduces a new polynomial soft threshold technique for denoising through wavelet shrinkage. The polynomial soft threshold technique is able to represent a wide class of polynomial behaviors, including classical soft thresholding.

This paper addresses the issue of oblivious robust watermarking, within the framework of colour still image database protection. We present an original method which complies with all the requirements nowadays imposed to watermarking applications: robustness (e.g. low-pass filtering, print & scan, StirMark), transparency (both quality and fidelity), low probability of false alarm, obliviousness and multiple bit recovering. The mark is generated from a 64 bit message (be it a logo, a serial number, etc.) by means of a Spread Spectrum technique and is embedded into DWT (Discrete Wavelet Transform) domain, into certain low frequency coefficients, selected according to the hierarchy of their absolute values. The best results were provided by the (9,7) bi-orthogonal transform. The experiments were carried out on 1200 image sequences, each of them of 32 images. Note that these sequences represented several types of images: natural, synthetic, medical, etc. and each time we obtained the same good results. These results are compared with those we already obtained for the DCT domain, the differences being pointed out and discussed.

This paper proposes a comparative study of 3D and 2D/3D shape descriptors (SDs) for indexing and retrieval of 3D mesh models. Seven state of the art SDs are considered and compared, among which five are 3D (Optimized 3D Hough Descriptor - O3DHD, Extended Gaussian Images - EGIs, cords length and spherical angles histograms, random triangles histogram, MPEG-7 3D shape spectrum descriptor - 3DSSD), and two 2D/3D, based on the MPEG-7 2D SDs (Contour Scale Space- CSS, and Angular Radial Transform - ART). A low complexity vector quantized (VQ) OH3DD is also proposed and considered for this comparison. Experimental results were carried out upon the categorized MPEG-7 3D test database. By computing Bull-Eye Score (BES) and First Tier (FT) criteria, it is objectively established that the O3DHD (even in its VQ version) outperforms (BES = 81% or 79%).all other SDs. The 2D/3D CSS-based descriptor exhibits a highly discriminant behavior (BES = 74%) outperforming the other both 3D and 2D/3D approaches. Apply to the industrial framework of the RNRT SEMANTIC-3D Project, the O3DHD demonstrated its relevance together with its scalability and robustness properties.

A variation of the matched spatial filter (MSF) has been developed and tested. It is derived by a power-series expansion of the ideal MSF that, in the absence of background noise, produces a Dirac delta function peak at the detection location on the correlation plane. The motivation for the approximation is to design an MSF that produces a narrow correlation peak with reduced susceptibility to noise amplification in the filtering process. This new filter, which we will call the complement MSF, includes the intrinsic phase information of the reference signal and a magnitude term that can be truncated at any order. Experimental results show that the complement MSF produces correlation peaks that can be controlled by varying the order of approximation, and that these peaks may improve discrimination over classical matched filtering methods such as the familiar cross correlation. In addition, we have exploited the familiar Wiener-Helstrom method for inverse filtering to blend both classical MSFs and complement MSFs of different order for imaging scenarios where the noise power spectrum is known or can be estimated. Preliminary outputs using this technique have shown sharper correlation peaks and better noise floor suppression than yielded by implementing the blended components individually.

This work has been carried out within the framework of the industrial project, so-called TOON, supported by the French government. TOON aims at developing tools for automating the traditional 2D cartoon content production. This paper presents preliminary results of the TOON platform. The proposed methodology concerns the issues of 2D/3D reconstruction from a limited number of drawn projections, and 2D/3D manipulation/deformation/refinement of virtual characters. Specifically, we show that the NURBS-based modeling approach developed here offers a well-suited framework for generating deformable 3D virtual characters from incomplete 2D information. Furthermore, crucial functionalities such as animation and non-rigid deformation can be also efficiently handled and solved. Note that user interaction is enabled exclusively in 2D by achieving a multiview constraint specification method. This is fully consistent and compliant with the cartoon creator traditional practice and makes it possible to avoid the use of 3D modeling software packages which are generally complex to manipulate.

This paper presents a novel algorithm for locating pupils in a portrait image for ID card application. The proposed algorithm composed of three steps; skin detection, eye detection, and pupil detection. Skin detection for reducing the region of interest employs three modified single Gaussian skin models. In the second step, candidates of horizontal and vertical eye locations are found by utilizing amount of deviation in R channel with an image that is cropped by skin detection. A small block centered at obtained coarse location is then further processed in pupil detection to find a precise pupil location. This step involves Pupil Index that measures the characteristics of pupil. If more than two locations are competing, ratios of Pupil Index and geometry rules are involved to select pupil locations. Experiments show that the algorithm successfully locates pupils. However, more works may need to be done on images that are rotated and/or tilted to a high degree.

Most of today's robot vehicles are equipped with omnidirectional sensors which provide surround awareness and easier navigation. Due to the persistence of the appearance in omnidirectional images, many global navigation or formation control tasks, instead of using landmarks or fiducials, they need only reference images of target positions or objects. In this paper, we study the problem of template matching in spherical images. The natural transformation of a pattern on the sphere is a 3D rotation and template matching is the localization of a target in any orientation given by a reference image. Unfortunately, the support of the template is space variant on the Euler angle parameterization. Here we propose a new method which matches the gradients of the image and the template, with space-invariant operation. Using properties of the angular momentum, we have proved in fact that the gradient correlation can be very easily computed by the 3D Inverse Fourier Transform of a linear combination of spherical harmonics. An exhaustive search localizes the maximum of this correlation. Experimental results on real data show a very accurate localization with a variety of targets. In future work, we plan to address targets appearing in different scales.

In this paper, a new face recognition system, GAYE, is presented. GAYE is a fully automatic system that detects and recognizes faces in cluttered scenes. The input of the system is any digitized image/image sequence that includes face/faces. The basic building blocks of the system are face detection, feature extraction and feature comparison. Face detection is based on skin color segmentation. For feature extraction, a novel approach is proposed that depends on the Gabor wavelet transform of the face image. By comparing facial feature vectors system finally makes a decision if the incoming person is recognized or not. Real time system tests show that GAYE achieves a recognition ratio over %90.

Dust detection and control in real time, represent one of the most challenging problem in all those environments where fine and ultrafine airborne particulate solids products are present. The presence of such products can be linked to several factors, often directly related and influenced by the working-production actions performed. Independently from the causes generating dust, airborne contaminants are an occupational problem of increasing interest as they are related to a wide number of diseases. In particular, airborne dusts are well known to be associated with several classical occupational lung diseases, such as the pneumoconiosis, especially at high levels of exposure. Nowadays there is also an increasing interest in other dust related diseases, from the most serious as cancer and asthma, to those related with allergies or irritation and other illnesses, also occurring at lower levels of exposure. Among the different critical factors influencing health risk for airborne dust exposure, mainly four have to be considered, that is: i) nature of the dust resulting from working in terms of presence of specific poisoning material, i.e. free silica, and morphological and morphometrical attributes of particulates constituting airborne dust; ii) size of the particles, iii) duration of exposure time and, finally, iv) airborne dust concentration in the breathing zone where the worker performs his activity. A correct dust detection is not easy, especially if some of the previous mentioned factors, have to be detected and quantified in real time in order to define specific “on-line” control actions aimed to reduce the level of the exposure to dust of the workers, as for example: i) modification of aspirating devices operating condition, change of filtering cleaning sequence, etc. . The more severe are the environmental conditions, in terms of dust presence (in quantity and quality) more difficult is to utilize efficient sampling devices. Detection devices, in fact, tend to become “blind” to dust as dust presence increases, on the other hand severe dust production conditions is exactly the case where control strategies have to be applied to realize safer conditions for the workers. In this paper the possibility to utilize a new logic to perform an “on-line” airborne dust sampling and analysis utilizing imaging is described with particular reference to dusts flowing in a duct after the caption and before their abatement by classical mechanical filtering. The study was particularly addressed to define, design and implement a logic able to extract those parameters affecting airborne dust behavior with respect to its efficient abatement. All dust sampling was performed directly in an industrial plant where tests were carried out in a controlled environment.

The Internet is a continuously expanding source of multimedia content and information. There are many products in development to search, retrieve, and understand multimedia content. But most of the current image search/retrieval engines, rely on a image database manually pre-indexed with keywords. Computers are still powerless to understand the semantic meaning of still or animated image content. Piria (Program for the Indexing and Research of Images by Affinity), the search engine we have developed brings this possibility closer to reality. Piria is a novel search engine that uses the query by example method. A user query is submitted to the system, which then returns a list of images ranked by similarity, obtained by a metric distance that operates on every indexed image signature. These indexed images are compared according to several different classifiers, not only Keywords, but also Form, Color and Texture, taking into account geometric transformations and variance like rotation, symmetry, mirroring, etc. Form - Edges extracted by an efficient segmentation algorithm. Color - Histogram, semantic color segmentation and spatial color relationship. Texture - Texture wavelets and local edge patterns. If required, Piria is also able to fuse results from multiple classifiers with a new classification of index categories: Single Indexer Single Call (SISC), Single Indexer Multiple Call (SIMC), Multiple Indexers Single Call (MISC) or Multiple Indexers Multiple Call (MIMC). Commercial and industrial applications will be explored and discussed as well as current and future development.

We present a generalized approach to dynamically incorporate high level knowledge into a cooperative intelligent image analysis framework. We developed this framework in our laboratory to provide a uniform interface to develop intelligent image analysis tools as well as to provide infrastructure facilities required by these tools in order to work cooperatively for accomplishing complex image analysis task goals. This framework is able to automatically generate processing plans which accomplish user defined image analysis task goals. The approach that we propose in this paper provides a flexible interface to develop the expertise of `image processing' tools. We provide two ways to develop this knowledge: 1) by taking feedback from an image processing expert about processing plans generated by the system; and 2) by accepting a processing plan which accomplishes a particular task from an expert user, and then extracting the high level knowledge encapsulated in this plan. The generalized nature of our approach allows each individual tool to use machine learning algorithms of its own interest in improving the knowledge-base. Preliminary results that we obtained from this work demonstrates the success of our approach.

Postural ability can be evaluated through the analysis of body oscillations, by estimating the displacements of selected sets of body segments. The analysis of human movement is generally achieved through the exploitation of stereophotogrammetric systems that rely on the use of markers. Marker systems show a high cost and patient settings which can be uncomfortable. On the other hand, the use of force platform has some disadvantages: the acquisition of dynamics data permits to estimate only the body oscillations as a whole, without any information about individual body segment movements. Some of these drawbacks can be overcome by the use of video systems, applying a marker-free sub-pixel algorithm. In this paper, a novel method to evaluate balance strategies that utilises commercial available systems and applies methods for feature extraction and image processing algorithms is presented.

Focus of attention is often attributed to biological vision system where the entire field of view is first monitored and then the attention is focused to the object of interest. We propose using a similar approach for object recognition in a color image sequence. The intention is to locate an object based on a prior motive, concentrate on the detected object so that the imaging device can be guided toward it. We use the abilities of the intelligent image analysis framework developed in our laboratory to generate an algorithm dynamically to detect the particular type of object based on the user's object description. The proposed method uses color clustering along with segmentation. The segmented image with labeled regions is used to calculate the shape descriptor parameters. These and the color information are matched with the input description. Gaze is then controlled by issuing camera movement commands as appropriate. We present some preliminary results that demonstrate the success of this approach.

In the framework of computer-aided diagnosis, pulmonary airway investigation based on multi-detector computerized tomography (MDCT) requires the development of specific tools for data interaction and analysis. The 3D segmentation of the bronchial tree provides radiologists with appropriate examination modalities such as CT bronchography, for a global analysis, or virtual endoscopy, for a local endoluminal diagnosis. Focusing on the latter modality, this paper proposes a set of advanced navigation and investigation tools based on the automatic extraction of the central axis (CA) of the 3D segmented airways. In the case of complex branching structures, such as the bronchial tree, the automatic CA computation is a challenging problem raising several difficulties related to geometry and topology preservation. In this respect, an original approach is presented, combining 3D distance map information and geodesic front propagation in order to accurately detect branching points and to preserve the original 3D topology of the airways, irrespective to both caliber variability with the bronchial order and to bronchial wall irregularities. The CA information is represented as a multi-valued and hierarchic tree structure, making possible automatic trajectory computation between two given points, bronchial caliber estimation in the plane orthogonal to the bronchus axis at a given location, branch indexation, and so on. These applications are illustrated on clinical data including both normal and pathological airway morphologies.

Computational mathematical morphology (CMM) is a nonlinear filter representation particularly amenable to real-time image processing. In the state of the art implementation each pixel value in a windowed observation is indexed into a separate lookup table to retrieve a set of bit vectors. Each bit in the vector corresponds to a basis element in the CMM filter representation. All retrieved bit vectors are "anded" together to produce a bit vector with a unique nonzero bit. The position of that bit corresponds to a basis element containing the observation and it used to look up a filter value in a table. The number of stored bit vectors is a linear function of the image or signal bit depth. We present an architecture for CMM implementation that uses a minimal number of bit vectors and required memory is less sensitive to bit depth. In the proposed architecture, basis elements are projected to subspaces and only bit vectors unique to each subspace are stored. With the addition of an intermediate lookup table to map observations to unique bit vectors, filter memory is greatly reduced. Simulations show that the architecture provides an advantage for random tessellations of the observation space. A 50% memory savings is shown for a practical application to digital darkness control in electronic printing.

This research is generally divided into two phases: the first phase deals with background image generation and vehicle detection, the second phase deals with vehicle tracking and video handoff. In the first phase we view the image as a mixture of three data distributions: vehicle, background and shadow. Thus the problem is modeled as a mixture of Gaussian problem and our goal is to separate the background data from other data distributions. We proposed a median model and an improved median model to separate the background data from mixture data and to generate background reference images. In median model we keep track of deviation between the median and its neighbors in a reordered pixel sequence. When sample size is big enough, the reordered pixel sequence is in what we called balanced-median model. This model is indicated by a very small deviation value. In this case the median of the pixel sequence falls in background set and could be used for background estimation. When sample size is not big enough, the reordered pixel sequence is in what we called shifted-median model. This model is indicated by a much bigger deviation value. In this case the median falls out of background set and are excluded for background estimation. This median model has an impressive performance to handle slow moving or even stationary vehicles. But the time complexity is still expensive for real time image processing. The improved median model is proposed to reduce the time complexity to a reasonable level. In improved median model, we take samples in a bigger time interval to make it capable of dealing with slow moving and stationary vehicles. The sample size from experimentation is obtained as a small constant value between 5 and 20. This small sample constant size could dramatically reduce the time complexity. As a complementary to this improved median model, a mask-classified updating method is introduced to update the background image in a short term and only classified background pixels are being used for updating. Threshold, erosion, dilation and connected components labeling are used for noise removing and object labeling. After the first phase, the vehicle information is separated from image and input to the second phase for video hand-off and vehicle tracking. In the second phase, the weighted intensity information and shape information for each vehicle is scored and minimum-distance classification method is used for vehicle match. More than 400 vehicles are tested. An overall detection rate of 100% and tracking rate of 74% are obtained in this system.

There are numerous applications for image fusion, some of which include medical imaging, remote sensing, nighttime operations and multi-spectral imaging. In general, the discrete wavelet transform (DWT) and various pyramids (such as Laplacian, ratio, contrast, gradient and morphological pyramids) are the most common and effective methods. For quantitative evaluation of the quality of fused imagery, the root mean square error (RMSE) is the most suitable measure of quality if there is a “ground truth” image available; otherwise, the entropy, spatial frequency or image quality index of the input images and the fused images can be calculated and compared. Here, after analyzing the pyramids’ performance with the four measures mentioned, an advanced wavelet transform (aDWT) method that incorporates principal component analysis (PCA) and morphological processing into a regular DWT fusion algorithm is presented. Specifically, at each scale of the wavelet transformed images, a principle vector was derived from two input images and then applied to two of the images’ approximation coefficients (i.e., they were fused by using the principal eigenvector). For the detail coefficients (i.e., three quarters of the coefficients), the larger absolute values were chosen and subjected to a neighborhood morphological processing procedure which served to verify the selected pixels by using a “filling” and “cleaning” operation (this operation filled or removed isolated pixels in a 3-by-3 local region). The fusion performance of the advanced DWT (aDWT) method proposed here was compared with six other common methods, and, based on the four quantitative measures, was found to perform the best when tested on the four input image types. Since the different image sources used here varied with respect to intensity, contrast, noise, and intrinsic characteristics, the aDWT is a promising image fusion procedure for inhomogeneous imagery.

Most remote sensing data-sets contain a limiting number of independent spatial and spectral measurements, beyond which no effective increase in information is achieved. This paper presents a Physically Motivated Correlation Formalism (PMCF) ,which places both Spatial and Spectral data on an equivalent mathematical footing in the context of a specific Kernel, such that, optimal combinations of independent data can be selected from the entire Hypercube via the method of "Correlation Moments". We present an experimental and computational analysis of Hyperspectral data sets using the Michigan Tech VFTHSI [Visible Fourier Transform Hyperspectral Imager] based on a Sagnac Interferometer, adjusted to obtain high SNR levels. The captured Signal Interferograms of different targets - aerial snaps of Houghton and lab-based data (white light , He-Ne laser , discharge tube sources) with the provision of customized scan of targets with the same exposures are processed using inverse imaging transformations and filtering techniques to obtain the Spectral profiles and generate Hypercubes to compute Spectral/Spatial/Cross Moments. PMCF answers the question of how optimally the entire hypercube should be sampled and finds how many spatial-spectral pixels are required for a particular target recognition.

A new concept of weighted thresholding is considered and a new set-theoretical representation of signals and images is described, that can be used for design of new nonlinear and morphological filters. Such representation maps many operations of nonbinary signal and image processing to the union of simple operations over the binary signals. The weighted thresholding is invariant under the morphological transformation, including such basic operations as the erosion and dilation. The main idea of using the weighted thresholding is in the choice of a special few levels of thresholding on which we can process the signals and images. We focus on the arithmetical weighted thresholding, but other thresholding, including the geometrical, probability-based, and the so-called Fibonacci series based thresholding, are also considered. Properties of these kinds of thresholding are described. Experimental results show that the weighted thresholding is very promising and can be used for many applications, such as image enhancement and edge detection.

One of the most common tasks in image processing is to change the resolution of a picture. In this paper we present a new nonlinear method for interpolating digital images, which is particulary effective in the rendition of edges in natural and synthetic input. The algorithm is spatial variant and applies the warped distance (WaDi) concept, generalizing the technique to a two dimensional problem, which requires a non-separable approach. It consists of three separate stages. First of all the original image is analyzed to detect its local gradient characteristics; then edge asymmetry is computed at each output pixel position according to the WaDi technique, and it is compared to a reference sigmoidal edge; the local edge asymmetry straightforwardly determines the warping factor which is applied to the bi-dimensional space of the image; eventually the actual interpolation is performed applying a conventional interpolator such as the linear or bicubic ones. The resulting interpolation method gives an output which does not present the usual blurring typical of images processed with linear interpolators and at the same time preserves the regularity of resized edges avoiding jagging artifacts. Moreover, the method adapts for zooming by a rational scaling factor. The paper is organized as follows. In the first section we introduce the problem of zooming digital images; the second section describes the state of the art; we continue describing the proposed method; then we propose a possible extension of the method to process color images; we end showing some examples of images interpolated with our method, and comparing these results with what can be obtained zooming the same input with other interpolators.

We propose a novel method for motion analysis in video sequences. It extends the co-occurrence matrix concept for texture analysis to the temporal domain. The approach proved to be versatile in the sense of targeting different motion analysis tasks. An application of the method is in the compact representation of video sequences, in particular temporal texture patterns.

Oriented patterns in an image often carry important information about the scene represented. Rao and Jain developed a technique to analyze images with oriented texture using phase portraits, where the parameters of a planar first-order phase portrait are locally estimated using a nonlinear least-squares algorithm. The method gives accurate results, but is computationally expensive. Shu and Jain proposed a faster linear method for the estimation of the parameters of the phase portrait. However, their formulation leads to the minimization of a different error measure, which is not as robust as the nonlinear least-squares procedure in the presence of noise, and also makes the implicit assumption that the orientation field was truly generated by a phase portrait model (with an extra weighting factor to compensate for noise sensitivity). We propose a new derivation of Shu and Jain's linear estimator that leads to similar estimation equations, while making explicit the nature of the error measure. Our procedure includes an iterative scheme, of which Shu and Jain's linear estimator is a particular case. We show that our estimator is more robust to noise than Shu and Jain's linear estimator.

Particle filtering is being investigated extensively due to its important feature of target tracking based on nonlinear and non-Gaussian model. It tracks a trajectory with a known model at a given time. It means that particle filter tracks an arbitrary trajectory only if the time instant when trajectory switches from one model to another model is known apriori. Because of this reason particle filter is not able to track any arbitrary trajectory where transition from one model to another model is not known. For real world application, trajectory is always random in nature and may follow more than one model. Another problem with multiple trajectories tracking using particle filter is the data association, i.e. observation to track fusion. In this paper we propose a novel method, which overcomes the above problems. In a proposed method an interacting multiple model based approach is used along with particle filtering, which automates the model selection process for tracking an arbitrary trajectory. We have utilized nearest neighbor (NN) method for data association, which is fast and easy to implement.

Comparing the output of a physics simulation with an experiment is often done by visually comparing the two outputs. In order to determine which simulation is a closer match to the experiment, more quantitative measures are needed. This paper describes our early experiences with this problem by considering the slightly simpler problem of finding objects in a image that are similar to a given query object. Focusing on a dataset from a fluid mixing problem, we report on our experiments using classification techniques from machine learning to retrieve the objects of interest in the simulation data. The early results reported in this paper suggest that machine learning techniques can retrieve more objects that are similar to the query than distance-based similarity methods.

In this paper, we propose a novel technique for blind image restoration and resolution enhancement based on radial basis function (RBF) neural network. The RBF network gives a solution of the regularization problem often seen in function estimation with certain standard smoothness functional used as stabilizers. A RBF network model is designed to represent the observed image. In this model, the number and distribution of the centers (which are set to the pixels of the observed image) are fixed. In addition, network output is set to the observed image pixel gray scale value. The RBF plays a role of point spread function. The technique can also be applied to image resolution enhancement by generating an interpolated image from the low resolution version. Experimental results show that the learning algorithm can effectively estimate the model parameters and the established neural network model has a high fidelity in representing an image. It is believed that the proposed neural network model provides a valuable tool for image restoration and resolution enhancement and holds promises to improve the quality and efficiency of image processing.

A robust version of Lee local statistic filter able to effectively suppress the mixed multiplicative and impulse noise in images is proposed. The performance of the proposed modification is studied for a set of test images, several values of multiplicative noise variance, Gaussian and Rayleigh probability density functions of speckle, and different characteris-tics of impulse noise. The advantages of the designed filter in comparison to the conventional Lee local statistic filter and some other filters able to cope with mixed multiplicative+impulse noise are demonstrated.

The paper is devoted to design, fast implementation and applications of a family of 8-points integer orthogonal transforms based on a parametric matrix. A unified algorithm for their efficient computations is developed. Derived fast transforms have close coding gain performance to the optimal Karhunen-Loeve transform for the first order Markov process. Among them are also such that closely approximate the DCT-II and, at the same time, have a larger coding gain. For a particular set of parameters, integer transforms with reduced computational complexity are obtained. The comparative analysis of these transforms with the DCT-II in the framework of image denoising and video coding is performed.

We present a cellular pulse coupled neural network with adaptive weights and its analog VLSI implementation. The neural network operates on a scalar image feature, such as grey scale or the output of a spatial filter. It detects segments and marks them with synchronous pulses of the corresponding neurons. The network consists of integrate-and-fire neurons, which are coupled to their nearest neighbors via adaptive synaptic weights. Adaptation follows either one of two empirical rules. Both rules lead to spike grouping in wave like patterns. This synchronous activity binds groups of neurons and labels the corresponding image segments. Applications of the network also include feature preserving noise removal, image smoothing, and detection of bright and dark spots. The adaptation rules are insensitive for parameter deviations, mismatch and non-ideal approximation of the implied functions. That makes an analog VLSI implementation feasible. Simulations showed no significant differences in the synchronization properties between networks using the ideal adaptation rules and networks resembling implementation properties such as randomly distributed parameters and roughly implemented adaptation functions. A prototype is currently being designed and fabricated using an Infineon 130nm technology. It comprises a 128 × 128 neuron array, analog image memory, and an address event representation pulse output.

Satellite image segmentation is an important task to generate classification maps. Land areas are classified and clustered into groups of similar land cover or land use by segmentation of satellite images. It may be broad classification such as urban, forested, open fields and water or may be more specific such as differentiating corn, soybean, beet and wheat fields. One of the most important among them is partitioning the urban area to different regions. On the other hand Multi-Channel filtering is used widely for texture segmentation by many researchers. This paper describes a texture segmentation algorithm to segment satellite images using Gabor filter bank and neural networks. In the proposed method feature vectors are extracted by multi-channel decomposition. The spatial/spatial-frequency features of the input satellite image are extracted by optimized Gabor filter bank. Some important considerations about filter parameters, filter bank coverage in frequency domain and the reduction of feature dimensions are discussed. A competitive network is trained to extract the best features and to reduce the feature dimension. Eventually a Multi-Layer Perceptron (MLP) is employed to accomplish the segmentation task. Our MLP uses the sigmoid transfer function in all layers and during the training, random selected feature vectors are assigned to proper classes. After MLP is trained the optimized extracted features are classified into sections according to the textured land cover regions.

In this paper, we propose a new segmentation method aimed at separating the moving objects from the background in a generic video sequence using Cellular Neural Networks (CNN). This task may be accomplished to support the functionalities foreseen by new multimedia scenarios, and in particular the content-based functionalities focused by the MPEG-4 activity. Extraction of motion information from video series is very power consuming, the proposed scheme extracts moving objects based on both motion and spatial information. Initially, a symmetrical inter-frame difference is performed on a group of gray image, so the approximate area of the video object was presented, then this area can be divided into some flat zones with uninterrupted grey scale information. Finally some zones are merged and forming the object according to a certain rule, others are discarded. It is the case of stationary background hereinbefore, in the case of moving, we will do some motion estimation at first. For the good of laborsaving, some work will be realized by CNN,. At the end of this paper, some typical results obtained on MPEG-4 sequences are here shown, in order to illustrate the segmentation algorithm performance using Aladdin V1.3 simulator system.

An unavoidable problem of metal structures is their exposure to rust degradation during their operational life. Thus, the surfaces need to be assessed in order to avoid potential catastrophes. There is considerable interest in the use of patch repair strategies which minimize the project costs. However, to operate such strategies with confidence in the long useful life of the repair, it is essential that the condition of the existing coatings and the steel substrate can be accurately quantified and classified. This paper describes the application of fuzzy set theory for steel surfaces classification according to the steel rust time. We propose a semi-automatic technique to obtain image clustering using the Fuzzy C-means (FCM) algorithm and we analyze two kinds of data to study the classification performance. Firstly, we investigate the use of raw images’ pixels without any pre-processing methods and neighborhood pixels. Secondly, we apply Gaussian noise to the images with different standard deviation to study the FCM method tolerance to Gaussian noise. The noisy images simulate the possible perturbations of the images due to the weather or rust deposits in the steel surfaces during typical on-site acquisition procedures

The paper introduces the approach to the 2D automated content-based object recognition utilizing the hybrid evolutionary algorithm (HEA), self-organizing network (SON), and response analysis. When the object is distorted and spatially misregistered, the recognition system has to solve a nonlinear global search problem, i.e. find simultaneously the global positioning of the object and the parameters of its local distortion. The task is accomplished with the HEA using the operators of selection and recombination for the global search, and the accelerated Downhill simplex method for the local search. The algorithm minimizes the fitness formulated as the normalized least squared difference between the images of the scene and the object, and utilizes image local response. The response adequately captures the dynamics of the image transformation, which makes it particularly well suited for the evolutionary search. The response matrix of the object is evaluated and presented to a SON. The weights are computed during the iterative learning process. The resulting adaptive response map serves as the footprint of the object. The evolutionary procedure identifies potential matches for the object based on the response matrix. The local refining procedure uses the response map to accelerate local search in the vicinity of the potential optimal solution.

Data association and model selection are important factors for tracking multiple targets in a dense clutter environment without using apriori information about the target dynamic. We propose a neural network based tracking algorithm, incorporating interacting multiple model to track both maneuvering and non-maneuvering targets simultaneously in the presence of dense clutter. For data association, we use the Expectation-Maximization (EM) algorithm and Hopfield network to evaluate assignment weights. All validated measurements are used to update the target state and hence, it avoids the uncertainty about the origin of the measurements. In the proposed approach the data association process is defined to incorporate multiple models for target dynamics and probability density function (pdf) of an observed data given target state and measurement association, is treated as a mixture pdf. This allows to combine the likelihood of a measurement due to each model, and consequently, it is possible to track any arbitrary trajectory in the presence of dense clutter.

In artificial neural networks (ANN) individual nodes are used as processing units that perform simple computations. The computations can be performed based on unsupervised or supervised learning schemes. One type of learning scheme is a competitive unsupervised approach. In the competitive approach different nodes compete to become the "winner(s)", representing the highest activity level. In a large array opto-electronic system the competitive dynamics is not restricted to single elements but is some specially distributed structure. This interaction is typical for partially distributed nonlinear systems with complex behavior but may be unusual behavior in other systems with large arrays of elements for example some ANN. With an opto-electronic system it may be possible to consider new dynamics and more complex behavior for systems with large arrays. A different approach for parallel high resolution information processing that potentially goes beyond processing large numbers of neurons or elements is considered. NN has been successful in processing low-resolution images. Hopefully opto-electronic systems can generate similar mechanisms seen in NN such as cooperation and competition. Perhaps different self-organizing structures or patterns generated by these systems have some features similar to competition and cooperation. These types of structure or pattern interactions can be possible building blocks for more robust computational processes.

The if-and-only-if (IFF) condition that a set of M analog-to-digital vector-mapping relations can be learned by a one-layered-feed-forward neural network (OLNN) is that all the input analog vectors dichotomized by the i-th output bit must be positively, linearly independent, or PLI. If they are not PLI, then the OLNN just cannot learn no matter what learning rules is employed because the solution of the connection matrix does not exist mathematically. However, in this case, one can still design a parallel-cascaded, two-layered, perceptron (PCTLP) to acheive this general mapping goal. The design principle of this "universal" neural network is derived from the major mathematical properties of the PLI theory - changing the output bits of the dependent relations existing among the dichotomized input vectors to make the PLD relations PLI. Then with a vector concatenation technique, the required mapping can still be learned by this PCTLP system with very high efficiency. This paper will report in detail the mathematical derivation of the general design principle and the design procedures of the PCTLP neural network system. It then will be verified in general by a practical numerical example.

Deinterlacing is the conversion process from the interlaced scan to progressive one. While many previous algorithms that are based on weighted-sum cause blurring in edge region, deinterlacing using neural network can reduce the blurring through recovering of high frequency component by learning process, and is found robust to noise. In proposed algorithm, input image is divided into edge and smooth region, and then, to each region, one neural network is assigned. Through this process, each neural network learns only patterns that are similar, therefore it makes learning more effective and estimation more accurate. But even within each region, there are various patterns such as long edge and texture in edge region. To solve this problem, modular neural network is proposed. In proposed modular neural network, two modules are combined in output node. One is for low frequency feature of local area of input image, and the other is for high frequency feature. With this structure, each modular neural network can learn different patterns with compensating for drawback of counterpart. Therefore it can adapt to various patterns within each region effectively. In simulation, the proposed algorithm shows better performance compared with conventional deinterlacing methods and single neural network method.

This paper proposes a new fractal image encoder using a SOFM neural network based classifier and also an improved isometric transformation, to reduce the encoding time. Here the sizes of a domain block and range block are 8x8 pixels and 4x4 pixels, respectively. Block is classified into one of four patterns, based on the variation of intensities of the pixels in the block: flat where it is very low, middle where it is small, vertical/horizontal where there exists a vertical or horizontal edge, diagonal where there exists a diagonal edge. The SOFM neural network memorizes these patterns by competitive learning where the weights on the connections are determined by the Kohenen's learning rules. To reduce the searching time, the proposed algorithm searches domain blocks following a spiral trajectory starting from the block selected in the range and uses an improved isometric transformation which classifies the templates before comparison. The experimental results have shown that the proposed algorithm reduces the encoding speed by 50% on average while maintaining the same PSNR and bit rate, compared to the other's recent research results.

There exist various objects, such as pictures, music, texts, etc., around our environment. We have a view for these objects by looking, reading or listening. Our view is concerned with our behaviors deeply, and is very important to understand our behaviors. Therefore, we propose a method which acquires a view as a vector, and apply the vector to sequence generation. We focus on sequences of the data of which a user selects from a multimedia database containing pictures, music, movie, etc.. These data cannot be stereotyped because user's view for them changes by each user. Therefore, we represent the structure of the multimedia database as the vector representing user's view and the stereotyped vector, and acquire sequences containing the structure as elements. We demonstrate a city-sequence generation system which reflects user's intension as an application of sequence generation containing user's view. We apply the self-organizing map to this system to represent user's view.

Traditionally, an unsupervised classification divides all pixels within an image into a corresponding class pixel by pixel. The number of classes must be selected, but seldom is ascertainable with little information in advance. Moreover, spectral properties of specific informational classes change seasonally for satellite imagery. The relationships between informational classes and spectral classes are not always constant, and relationships defined for one image cannot be extended to others. Thus, the analyst has very limited or no control over the menu of classes and their specific identities. In this study, a Genetic Algorithm is adopted to interpret the cluster centers of an image and to reveal a suitable number of classes to overcome the disadvantage of unsupervised classification. A Genetic Algorithm is capable of dealing with a set of numerous data such as satellite imagery pixels. An optimization consequence of the image classification is introduced and carried out. Through an image process program developed in Mathlab, the GA unsupervised classifier was processed on several test images for validity and on SPOT satellite imagery. The classified SPOT image was compared with finer aerial photographs as a ground truth for the estimation of classification accuracy.

Non-linear techniques for denoising images and video are known to be superior to linear ones. In addition video denoising using spatio-temporal information is considered to be more efficient compared with the use of just temporal information in the presence of fast motion and low noise. Earlier, we introduced a 3-D extension of the K-nearest neighbor filter and have investigated its properties. In this paper we propose a new, motion- and detail-adaptive filter, which solves some of the potential drawbacks of the non-adaptive version: motion caused artifacts and the loss of fine details and texture. We also introduce a novel noise level estimation technique for automatic tuning of the noise-level dependent parameters. The results show that the adaptive K-nearest neighbor filter outperforms the none-adaptive one, as well as some other state-of-the-art spatio-temporal filters such as the 3D alpha-trimmed mean and the state-of-the-art rational filter by Ramponi from both a PSNR and visual quality point of view.

This paper presents an affine point-set and line invariant algorithm within a statistical framework, and its application to photo-identification of gray whales (Eschrichtius robustus). White patches (blotches) appearing on a gray whale's left and right flukes (the flattened broad paddle-like tail) constitute unique identifying features and have been used here for individual identification. The fluke area is extracted from a fluke image via the live-wire edge detection algorithm, followed by optimal thresholding of the fluke area to obtain the blotches. Affine point-set and line invariants of the blotch points are extracted based on three reference points, namely the left and right tips and the middle notch-like point on the fluke. A set of statistics is derived from the invariant values and used as the feature vector representing a database image. The database images are then ranked depending on the degree of similarity between a query and database feature vectors. The results show that the use of this algorithm leads to a reduction in the amount of manual search that is normally done by marine biologists.

In this paper, we describe a real time vehicle tracking using image processing techniques. The moving vehicles are segmented from the input image sequence using differential edge images. The vehicles are tracked using statistical invariant moments. The direction of vehicles is determined by the Hough transform. The direction of the vehicles is determined from the straight lines in the direction of the principal axes of the vehicles. The motion information is calculated from the displacement of the vehicles and the change of direction of vehicles in the consecutive frames. The algorithm is tested on different real time image sequences.

This paper describes a new shape matching using image projection features. The corners are located in the binary image using the radius vector function. The imaginary line joining any two corners is called the baseline, if the distance between those two corners is the maximum. The image is rotated to align the baseline with the reference axis. Horizontal and vertical projections of the rotated image are drawn. The projections are matched with the projections of the database images using the sorted normalized matching algorithm. The algorithm is tested on various test images.

An automatic locating algorithm is presented for typhoon center locating using cloud motion wind vectors derived from the satellite cloud images. The cloud motion wind vectors are obtained by implementing template matching to a pair of interrelated satellite cloud images with stated time interval. The template matching is a process to find the child image that corresponds to the given pattern image in an unknown pattern image. Three matching algorithms are compared. Namely, the absolute difference matching algorithm, the sequential similarity detection algorithm and the infrared cross-correlation coefficients matching algorithm. The third one is selected to acquire the set of cloud motion wind vectors duo to its desirable vector results. Aiming at the specific typhoon cloud image, two simplifications are processed in the course of acquiring the cloud motion wind vectors. According to meteorological analysis, typhoon center motion has two important characteristics: (1) The translation in the central area is great while the spin is feeble. (2) The center moving direction is compatible to that of the whole typhoon clouds. According to these characteristics, the algorithm for automatically locating the typhoon center can be depicted as follows: firstly pick up the vectors that compatible to the whole typhoon cloud motion vectors in the cloud motion wind vectors image, then find out the thickest area of the satisfied vectors, lastly process the thickest area with mathematical morphology until there exists only one pixel point. The locating result shows that the thought in the paper is good and can be a promising application in the typhoon center location field.

Independent component analysis (ICA) is a way to resolve signals into independent components based on the statistical characteristics of the signals. It is a method for factoring probability densities of measured signals into a set of densities that are as statistically independent as possible under the assumptions of a linear model. Electrical impedance tomography (EIT) is used to detect variations of the electric conductivity of the human body. Because there are variations of the conductivity distributions inside the body, EIT presents multi-channel data. In order to get all information contained in different location of tissue it is necessary to image the individual conductivity distribution. In this paper we consider to apply ICA to EIT on the signal subspace (individual conductivity distribution). Using ICA the signal subspace will then be decomposed into statistically independent components. The individual conductivity distribution can be reconstructed by the sensitivity theorem in this paper. Compute simulations show that the full information contained in the multi-conductivity distribution will be obtained by this method.

In the dental field, the 3D tooth model in which each tooth can be manipulated individually is an essential component for the simulation of orthodontic surgery and treatment. To reconstruct such a tooth model from CT slices, we need to define the accurate boundary of each tooth from CT slices. However, the global threshold method, which is commonly used in most existing 3D reconstruction systems, is not effective for the tooth segmentation in the CT image. In tooth CT slices, some teeth touch with other teeth and some are located inside of alveolar bone whose intensity is similar to that of teeth. In this paper, we propose an image segmentation algorithm based on B-spline curve fitting to produce smooth tooth regions from such CT slices. The proposed algorithm prevents the malfitting problem of the B-spline algorithm by providing accurate initial tooth boundary for the fitting process. This paper proposes an optimal threshold scheme using the intensity and shape information passed by previous slice for the initial boundary generation and an efficient B-spline fitting method based on genetic algorithm. The test result shows that the proposed method detects contour of the individual tooth successfully and can produce a smooth and accurate 3D tooth model for the simulation of orthodontic surgery and treatment.

A problem of surface imaging and parameter determination in an antenna far-field range by scanning the electromagnetic field intensity distribution in a plane combined with the imagined position of a test antenna aperture is considered. The signal at the receiver output is formed as a result of superposition in the antenna of the direct and reflected waves from the range surface. Exact and approximate mathematical models for description of an "Antenna - Range" system are offered and analysed. Model parameters are determining by statistical regression method. The system parameter dependence on model parameters is obtained. The offered method allows determining both the path length and direction of the reflected waves, the sites of the antenna range essential in reflection, as well as all system parameters, including the module of reflection coefficient. Processing of both modeled and measured signals, as well as surface imaging of a virtual and the real antenna ranges are carried out. The results may be used during operation of any existing antenna range, as well as in their design.

The background and foreground modeling is essential in tracking objects from the scenes taken by the stationary camera. We suggest a background model using moving histogram method. A moving histogram, which can be called pixel-wise approach, is time-dependent and can be regarded as a probability density function (pdf) of intensity in image sequence. This moving histogram is updated using image sequence from a stationary camera and is used to calculate the probability of which a pixel in incoming image belongs to background model. Pixels failed in entering into the background model can be candidates for foreground objects. These pixels are classified into foreground ones by comparing with other candidate pixels in different image frames. For pixel classification, our background process consists of queue memory which stores recently acquired images. The background process updates moving histogram for each (x, y) pixel and computes maximum frequency pixel value with low computation. After updating the moving histogram, the background process classifies each pixel as the moving pixel or the background pixel. The classification is difficult because of the slow change in background brightness, slow moving objects, clutters, and the shadow. We solve this problem heuristically. The moving histogram consists of several models (multi-modal, vehicle, background, shadow, clutter). We can compute the distance between the incoming pixel value and each model. And we use threshold with Euler numbers for foreground segmentation. The background and the segmentation process need small computation and can be adapted easily to real-time system.

Texture segmentation and analysis is an important aspect of pattern recognition and digital image processing. Previous approaches to texture analysis and segmentation perform multi-channel filtering by applying a set of filters to the image. In this paper we describe a texture segmentation algorithm based on multi-channel filtering that is optimized using diagonal high frequency residual. Gabor band pass filters with different radial spatial frequencies and different orientations have optimum resolution in time and frequency domain. The image is decomposed by a set of Gabor filters into a number of filtered images; each one contains variation of intensity on a sub-band frequency and orientation. The features extracted by Gabor filters have been applied for image segmentation and analysis. There are some important considerations about filter parameters and filter bank coverage in frequency domain. This filter bank does not completely cover the corners of the frequency domain along the diagonals. In our method we optimize the spatial implementation for the Gabor filter bank considering the diagonal high frequency residual. Segmentation is accomplished by a feedforward backpropagation multi-layer perceptron that is trained by optimized extracted features. After MLP is trained the input image is segmented and each pixel is assigned to the proper class.

An information fusion based fuzzy segmentation method applied to Magnetic Resonance Images (MRI) is proposed in this paper. It can automatically extract the normal and abnormal tissues of human brain from multispectral images such as T1-weighted, T2-weighted and Proton Density (PD) feature images. Fuzzy models of normal tissues corresponding to three MRI sequences images are derived from histogram according to a priori knowledge. Three different functions are chosen to calculate the fuzzy models of abnormal tissues. Then, the fuzzy features extracted by these fuzzy models are joined by a fuzzy relation operator which represents their fuzzy feature fusion. The final segmentation result is obtained by a fuzzy region growing based fuzzy decision rule. The experimental results of the proposed method are compared with the manually labeled segmentation by a neuroradiologist for abnormal tissues and with anatomic model of BrainWeb for normal tissues. The MRI images used in our experiment are imaged with a 1.5T GE for abnormal brain, with 3D MRI simulated brain database for normal brain by using an axial 3D IR T1-weighted (TI/TR/TE: 600/10/2), an axial FSE T2-weighted(TR/TE: 3500/102) and an axial FSE PD weighted (TR/TE: 3500/11). Based on 4 patients studied, the average probability of false detection of abnormal tissues is 5%. For the normal tissues, a false detection rate of 4% - 15% is obtained in images with 3% - 7% noise level. All of them show a good performance for our method.

In this contribution, we explore the best basis paradigm for in feature extraction. According to this paradigm, a library of bases is built and the best basis is found for a given signal class with respect to some cost measure. We aim at constructing a library of anisotropic bases that are suitable for the class of 2-D binarized character images. We consider two, a dyadic and a non-dyadic generalization scheme of the Haar wavelet packets that lead to anisotropic bases. For the non-dyadic case, generalized Fibonacci p-trees are used to derive the space division structure of the transform. Both schemes allow for an efficient O(NlogN) best basis search algorithm. The so built extended library of anisotropic Haar bases is used in the problem of optical character recognition. A special case, namely recognition of characters from very low resolution, noisy TV images is investigated. The best Haar basis found is then used in the feature extraction stage of a standard OCR system. We achieve very promising recognition rates for experimental databases of synthetic and real images separated into 59 classes.

In a previous study, a new adaptive method (AM) was developed to adjust the learning rate in artificial neural networks: the generalized no-decrease adaptive method (GNDAM). The GNDAM is fundamentally different from other traditional AMs. Instead of using the derivative sign of a given weight to adjust its learning rate, this AM is based on a trial and error heuristic where global learning rates are adjusted according to the error rates produced by two identical networks using different learning rates. This AM was developed to solve a particular task: the orientation detection of an image defined by texture (the texture task). This new task is also fundamentally different from other traditional ones since its data set is infinite, each pattern is a template used to generate stimuli that the network learns to classify. In the previous study, the GNDAM showed its strength over standard backpropagation for this particular task. The present study compares this new AM to other traditional AMs on the texture task and other benchmark tasks. The results showed that some AMs work well for some tasks while others work better for other tasks. However, all of them failed to achieve a good performance on all tasks.

For any modern 3D vision guided system, it is imperative to have complete range images for 3D model reconstruction. In practice, depth images obtained from a standard stereo camera can be error-prone with missing depth pixels. This paper proposes a method to augment the range data obtained from stereoscopy with model based image segmentation using planar patches. The method integrates both intensity and range data obtained from a standard stereo system. First, edges are extracted and linked to different segments from the intensity image with the embedded confidence edge detection technique and a general edge-linking algorithm respectively. Since edges are where disparity happens, most straight-line edges segmented from linked edges using a line-curvature extraction algorithm will have valid depth data. The planar patches are then defined by the straight-lines edge. With the knowledge of the planar structure, each depth missing pixels in the region is then determined by various 3-D line equations that pass through the pixel in the world coordinate system. Lastly, the range data in the world coordinate system is converted back into the image coordinate system using a pixel-to-pixel project algorithm. Result demonstrates the accuracy of method for filling up the missing depth in a region.

We develop a level set based region growing method for automatic partitioning of color images into segments. Previous attempts at image segmentation either suffer from requiring a priori information to initialize regions, being computationally complex, or fail to establish the color consistency and spatial connectivity at the same time. Here, we represent the segmentation problem as monotonic wave propagation in an absorbing medium with varying front speeds. We iteratively emit waves from the selected base points. At a base point, the local variance of the data reaches a minimum, which indicates the base point is a suitable representative of its local neighborhood. We determine local variance by applying a hierarchical gradient operator. The speed of the wave is determined by the color similarity of the point on the front to the current coverage of the wave, and by edge information. Thus, the wave advances in an anisotropic spatial-color space. The absorbing function acts as a stopping criterion of the wave front. We take advantage of fast marching methods to solve the Eikonal equation for finding the travel times of the waves. Besides, each region boundary is represented as a mixture of Gaussian models. This formulation enables segmentation of multi-modal color objects. Our method is superior to the linkage-based and snake-based region growing techniques since it prevents leakage and imposes compactness on the region without over-smoothing its boundary. Furthermore, we can deal with sharp corners and changes in topology. The automatic segmentation method is Eulerian, thus it is computationally efficient. Our experiments illustrate the robustness, accuracy, and effectiveness of the proposed method.

In order to provide an efficient way to processing with limited resources, we propose a novice wavelet coder that operates with little memory usage on the portable embedded system. In order to reduce redundancy in coding process caused by repetitive scanning of wave let coefficients, the proposed coder uses a 2-D significance coefficient array (SCA) which records the bit-level information of wavelet coefficients. The 2-D SCA improves memory usage and processing speed required for image coding because it can perform significance check and bit coding of coefficients simultaneously.

Many computer vision and image processing algorithms rely on the knowledge of the image noise variance as their input parameter. However, in practice, the distinction between noise and image features is not easy to draw. In this paper, image noise variance is estimated by a novel method employing rigorously derived polynomial masks. The method is based on the assumption that the image can be locally represented as a polynomial of the given degree and constitutes a generalization of some of previously proposed approaches.

We present the analysis and simulation results for some modifications of the vectorial color imaging procedures those use at the second stage of magnitude processing the different order statistics filters. The technique of non-parametric filtering is presented and investigated in this paper too. For unknown functional form of noise density estimated from the observations we use the gray scalar filters to provide the reference vectors needed to realize the calculations. The performances of the traditional order statistics algorithms such as, median, Vector Median, alfa-trimmed mean, Wilcoxon, other order statistics M KNN are analyzed in the paper. For comparison analysis of the color imaging we use the following criterions: MAE; PSNR; MCRE; NCD Numerous simulation results which characterize the impulsive noise suppression and fine detail preservation are presented in the paper using different test images) such as: Lena, Mandrill, Peppers, etc. (256x256, 24 bits, RGB space). The algorithms those demonstrated good performance results have been applied to process the video sequences: “Miss America”, “Flowers” and Foreman” corrupted by impulsive noise. The results of the simulations presented in the paper show differences in color imaging by mentioned filtering technique and help to choose the filter that can satisfy to several criterion at dependence on noise level value.