This PDF file contains the front matter associated with SPIE Proceedings Volume 6497, including the Title Page, Copyright information, Table of Contents, and the Conference Committee listing.

Canny edge detector is based both on local and global image analysis, present in the gradient computation and connectivity-related hysteresis thresholding, respectively. This contribution proposes a generalization of these ideas. Instead of the sole gradient magnitude, we consider several local statistics, to take into account how much texture is present around each pixel. This information is used in biologically inspired surround inhibition of texture. Global analysis is generalized by introducing a long range connectivity analysis. We demonstrate the effectiveness of our approach by extensive experimentation.

Object detection in images was conducted using a nonlinear means of improving signal to noise ratio termed "stochastic resonance" (SR). In a recent United States patent application, it was shown that arbitrarily large signal to noise ratio gains could be realized when a signal detection problem is cast within the context of a SR filter. Signal-to-noise ratio measures were investigated. For a binary object recognition task (friendly versus hostile), the method was implemented by perturbing the recognition algorithm and subsequently thresholding via a computer simulation. To fairly test the efficacy of the proposed algorithm, a unique database of images has been constructed by modifying two sample library objects by adjusting their brightness, contrast and relative size via commercial software to gradually compromise their saliency to identification. The key to the use of the SR method is to produce a small perturbation in the identification algorithm and then to threshold the results, thus improving the overall system's ability to discern objects. A background discussion of the SR method is presented. A standard test is proposed in which object identification algorithms could be fairly compared against each other with respect to their relative performance.

Corner matching is an important operation in digital image processing and computer vision where it is used for a range of applications including stereo vision and image registration. A number of corner similarity metrics have been developed to facilitate matching, however, any individual metric has a limited effectiveness depending on the content of images to be registered and the different types of distortions that may be present. This paper explores combining corner similarity metrics to produce more effective measures for corner matching. In particular the combination of two similarity metrics is investigated using experiments on a number of images exhibiting different types of transformations and distortions. The results suggest that a linear combination of different similarity metrics may produce more accurate and robust assessments of corner similarity.

This paper proposes a junction detection method that detects junctions as those points where edges join or intersect. The edges that form a junction are searched in a square neighbourhood, and the subtended angles among them are calculated by using edge orientations. Local edge orientation at a pixel is estimated by utilizing those edge points close to the pixel. Based on the subtended angles, the pixel is determined to be a junction candidate or not. Each actual junction is accurately localized by suppressing the candidates of non-minimum orientation difference. The proposed method analyzes real cases of extracted edges, and estimates the change of orientations of edge segments in digital fields. The experimental results show that the proposed algorithm can robustly detect junctions in digital images.

Conventional approach in single-chip digital cameras is a use of color filter arrays (CFA) in order to sample different spectral components. Demosaicing algorithms interpolate these data to complete red, green, and blue values for each image pixel, in order to produce an RGB image. In this paper we propose a novel demosaicing algorithm for the Bayer CFA. It is assumed that the initial estimates of color channels contain two additive components: the true values of color intensities and the errors. The errors are considered as an additive noise, and often called as a demosaicing noise, that has been removed. However, this noise is not white and strongly depends on a signal. Usually, the intensity of this noise is higher near edges of image details. We use spatially designed signal-adaptive filter to remove the noise. This filter is based on the local polynomial approximation (LPA) and the paradigm of the intersection of confidence intervals (ICI) applied for selection adaptively varying scales (window sizes) of LPA. The LPA-ICI technique is nonlinear and spatially-adaptive with respect to the smoothness and irregularities of the image. The efficiency of the proposed approach is demonstrated by simulation results.

A fundamental challenge in analyzing spatial patterns in images is the notion of scale. Texture based analysis typically characterizes spatial patterns only at the pixel level. Such small scale analysis usually fails to capture spatial patterns that occur over larger scales. This paper presents a novel solution, termed hierarchical texture motifs, to this texture-of-textures problem. Starting at the pixel level, spatial patterns are characterized using parametric statistical models and unsupervised learning. Higher levels in the hierarchy use the same analysis to characterize the motifs learned at the lower levels. This multi-level analysis is shown to outperform single-level analysis in classifying a standard set of image textures.

Dust, scratches or hair on originals (prints, slides or negatives) distinctly appear as light or dark artifacts on a scan. These unsightly artifacts have become a major consumer concern. This paper describes an algorithmic solution to the dust and scratch removal task. The solution is divided into two phases: a detection phase and a reconstruction phase. Some scanners have dedicated hardware to detect dust and scratch areas in the original. Without hardware assistance, dust and scratch removal algorithms generally resort to blurring, at the loss of image detail. We present an algorithmic alternative for dust and scratch detection that effectively differentiates between defects and image details. In addition we present reconstruction algorithms, that preserve image sharpness better than available alternatives. For detection we generate a detail-less image in which the defects are "erased". We compare properties of the luminance channel of the input image relative to the detailless image. For reconstruction of the defective areas we suggest both a fast small support algorithm and a large support algorithm, which is better able to mimic the existing image texture.

To understand a comprehensive atmospheric state, it is important to classify clouds in satellite images into appropriate classes. Many researches utilizing various features concerning the cloud texture have been reported in cloud classification. However, some clouds can not be classified uniquely only with the texture features. According to the knowledge of the experts, they classify the clouds in two stages. They firstly categorize the clouds into the provisional classes according to the brightnesses of the satellite images. They then classify each provisional class into the objective class based on the texture, shape and velocity of the cloud employing the meteorological knowledge about the time and location of the image. In this paper, we propose a novel method for the cloud classification that consists of two stages and utilizes cloud movement as human experts adopt. We firstly classify the clouds into 20 classes based on their brightnesses of the two-band spectral images. We then closely analyze the classes according to five features such as the brightnesses, deviations of brightness and cloud velocity estimated by varying window size adaptively. The experimental results are shown to verify the proposed method.

Our challenge was to develop a semi-automatic target detection algorithm to aid human operators in locating potential targets within images. In contrast to currently available methods, our approach is relatively insensitive to image brightness, image contrast and object orientation. Working on overlapping image blocks, we used a sliding difference method of histogram matching. Incrementally sliding the histograms of the known object template and the image region of interest (ROI) together, the sum of absolute histogram differences was calculated. The minimum of the resultant array was stored in the corresponding spatial position of response surface matrix. Local minima of the response surface suggest possible target locations. Because the template contrast will rarely perfectly match the contrast of the actual image contrast, which can be compromised by illumination conditions, background features, cloud cover, etc., we perform a random contrast manipulation, which we term 'wobble', on the template histogram. Our results have shown increased object detection with the combination of the sliding histogram difference and wobble.

With the increased emphasis on security and personal authentication, an accurate biometric-based authentication system has become a critical requirement in a variety of applications. Among different biometrics, authentication based on iris features has received a lot of attention since its introduction in 1992. The wavelet transform has been proposed by several researchers for extracting iris features for authentication. Although classical wavelets provide a good performance, they suffer from limited orientation selectivity. In this paper, we investigate the potentials of using the contourlet transform to represent the iris texture. A new iris representation and matching system based on contourlet transform is proposed. The contourlet transform not only shares the multiscale and localization properties of wavelets, but also has a higher degree of directionality and anisotropy. The proposed matching system is experimented in both verification and identification modes. Results have shown the significance of the new technique, especially in case of low quality iris images and highly security demanding applications.

In this paper,we study optical flow determination with Complex Sinusoidally Modulated Imaging (CSMI) using a novel imaging device, the correlation image sensor (CIS). The proposed method is based on a newly obtained relation what we call the optical flow identity (OFI) between intensity image and complex inusoidally-modulated image captured simultaneously by the CIS. This equation is complex-valued and the optical flow is a 2-dimensional vector for each pixel. Therefore, it is possible to compute the optical flow from one pixel value and its spatial gradient. Since the OFI does not involve time derivative, information on a single frame is sufficient. Moreover, the velocity limitation due to the spatio-temporal aliasing and approximate frame differentials used in conventionalt methods is avoided.

In the context of site-specific weed management by vision systems, an efficient image processing for a crop/weed discrimination is required in order to quantify the Weed Infestation Rate (WIR) in an image. This paper presents a modeling of crop field in presence of different Weed Infestation Rates and a set of simulated agronomic images is used to test and validate the effectiveness of a crop/weed discrimination algorithm. For instance, an algorithm has been implemented to firstly detect the crop rows in the field by the use of a Hough Transform and secondly to detect plant areas by a region based-segmentation on binary images. This image processing has been tested on virtual cereal fields of a large field of view with perspective effects. The vegetation in the virtual field is modeled by a sowing pattern for crop plants and the weed spatial distribution is modeled by either a Poisson process or a Neyman-Scott cluster process. For each simulated image, a comparison between the initial and the detected weed infestation rate allows us to assess the accuracy of the algorithm. This comparison demonstrates an accuracy of better than 80% is possible, despite that intrarow weeds can not be detected from this spatial method.

Robust image-based motion stabilization is developed to enable visual surveillance in the maritime domain. The algorithm developed is neither a dense registration method nor a traditional feature-based method, but rather it captures the best aspects of each of these approaches. It avoids feature tracking and so can handle large intra-frame motions, and at the same time it is robust to large lighting variations and moving clutter. It is thus well-suited for challenges in the maritime domain. Advantage is taken of the maritime environment including use of the horizon and shoreline, and fused data from an inexpensive inertial measurement unit. Results of real-time operation on an in-water buoy are presented.

Image enhancement is the task of applying certain alterations to an input image such as to obtain a more visually pleasing image. The alteration usually requires interpretation and feedback from a human evaluator of the output resulting image. Therefore, image enhancement is considered a difficult task when attempting to automate the analysis process and eliminate the human intervention. Furthermore, images that do not have uniform brightness pose a challenging problem for image enhancement systems. Different kinds of histogram equalization techniques have been employed for enhancing images that have overall improper illumination or are over/under exposed. However, these techniques perform poorly for images that contain various regions of improper illumination or improper exposure. In this paper, we introduce new human vision model based automatic image enhancement techniques, multi-histogram equalization as well as local and adaptive algorithms. These enhancement algorithms address the previously mentioned shortcomings. We present a comparison of our results against many current local and adaptive histogram equalization methods. Computer simulations are presented showing that the proposed algorithms outperform the other algorithms in two important areas. First, they have better performance, both in terms of subjective and objective evaluations, then that currently used algorithms on a series of poorly illuminated images as well as images with uniform and non-uniform illumination, and images with improper exposure. Second, they better adapt to local features in an image, in comparison to histogram equalization methods which treat the images globally.

Iris pattern matching is a key step in iris patter recognition system. This paper proposes a new approach for iris pattern matching using block-based spectral angle measure (SAM). The iris patterns in this paper are extracted using Log-Gabor wavelet method. The proposed SAM method will capture not only the phase information, but also the magnitude information. The block based method will enable the flexibility in the pattern matching step, which improves the accuracy and is more tolerable of the pattern dilation/constrain resulted from imperfect preprocessing. The preliminary experimental results show that the proposed method has an encouraging performance.

Video image processing based vehicle tracking and traffic monitoring provides several advantages over traditional approaches. One of the challenges lies in the robust segmentation and region tracking. The traditional approaches have assumed 1) the region of the object of interest is uniform and homogeneous; 2) adjacent regions should differ significantly. These are often not true for a vehicle object. In this paper, we propose a dynamic content based image segmentation method for vehicle tracking and traffic monitoring. Only initial lane information is needed for camera calibration. The system will automatically detect the direction of the traffic flow for vehicle detection and traffic monitoring. The preliminary experimental results show that this method is effective.

What visually distinguishes a painting from a photograph is often the absence of texture and the sharp edges: in many paintings, edges are sharper than in photographic images while textured areas contain less detail. Such artistic effects can be achieved by filters that smooth textured areas while preserving, or enhancing, edges and corners. However, not all edge preserving smoothers are suitable for artistic imaging. This study presents a generalization of the well know Kuwahara filter aimed at obtaining an artistic effect. Theoretical limitations of the Kuwahara filter are discussed and solved by the new nonlinear operator proposed here. Experimental results show that the proposed operator produces painting-like output images and is robust to corruption of the input image such as blurring. Comparison with existing techniques shows situations where traditional edge preserving smoothers that are commonly used for artistic imaging fail while our approach produces good results.

In this paper we propose a novel color demosaicing algorithm for noisy data. It is assumed that the data is given according to the Bayer pattern and corrupted by signal-dependant noise which is common for CCD and CMOS digital image sensors. Demosaicing algorithms are used to reconstruct missed red, green, and blue values to produce an RGB image. This is an interpolation problem usually called color filter array interpolation (CFAI). The conventional approach used in image restoration chains for the noisy raw sensor data exploits denoising and CFAI as two independent steps. The denoising step comes first and the CFAI is usually designed to perform on noiseless data. In this paper we propose to integrate the denoising and CFAI into one procedure. Firstly, we compute initial directional interpolated estimates of noisy color intensities. Afterward, these estimates are decorrelated and denoised by the special directional anisotropic adaptive filters. This approach is found to be efficient in order to attenuate both noise and interpolation errors. The exploited denoising technique is based on the local polynomial approximation (LPA). The adaptivity to data is provided by the multiple hypothesis testing called the intersection of confidence intervals (ICI) rule which is applied for adaptive selection of varying scales (window sizes) of LPA. We show the efficiency of the proposed approach in terms of both numerical and visual evaluation.

A practical impossibility of prediction of signs of DCT coefficients is generally accepted. Therefore each coded sign of DCT coefficients occupies usually 1 bit of memory in compressed data. At the same time data of all coded signs of DCT coefficients occupy about 20-25% of a compressed image. In this work we propose an effective approach to predict signs of DCT coefficients in block based image compression. For that, values of pixels of already coded/decoded neighbor blocks of the image are used. The approach consist two stages. At first, values of pixels of a row and a column which both are the nearest to already coded neighbor blocks are predicted by a context-based adaptive predictor. At second stage, these row and column are used for prediction of the signs of the DCT coefficients. Depending on complexity of an image proposed method allows to compress signs of DCT coefficients to 60-85% from their original size. It corresponds to increase of compression ratio of the entire image by 3-9% (or up to 0.5 dB improvement in PSNR).

The dynamic range of an image is defined as the ratio between the maximum and minimum luminance value it contains. This value in real images can be several thousands or even millions, whereas the dynamic range of consumer imaging devices rarely exceeds 100; therefore some processing is needed in order to display a high dynamic range image correctly. Global operators map each pixel individually with the same nonlinear function; local operators use spatially-variant functions in order to achieve a higher quality. The lower computational cost of global operators makes them attractive for real-time processing; the nonlinear mapping can however attenuate the image details. In this paper we define an expression which gives a quantitative measure of this artifact, and compare the performance of some commonly used operators. We show that a modified logarithm we propose has a satisfactory performance for a wide class of images, and has a theoretical justification based on some properties of the human visual system. We also introduce a method for the automatic tuning of the parameters of our system, based on the statistics of the input image. We finally compare our method with others proposed in the literature.

Classification of image segments on textures can be helpful for target recognition. Sometimes target cueing is performed before target recognition. Textures are sometimes used to cue an image processor of a potential region of interest. In certain imaging sensors, such as those used in synthetic aperture radar, textures may be abundant. The textures may be caused by the object material or speckle noise. Even speckle noise can create the illusion of texture, which must be compensated in image pre-processing. In this paper, we will discuss how to perform texture classification but constrain the number of wavelet packet node decomposition. The new approach performs a twochannel wavelet decomposition. Comparing the strength of each new subband with others at the same level of the wavelet packet determines when to stop further decomposition. This type of decomposition is performed recursively. Once the decompositions stop, the structure of the packet is stored in a data structure. Using the information from the data structure, dominating channels are extracted. These are defined as paths from the root of the packet to the leaf with the highest strengths. The list of dominating channels are used to train a learning vector quantization neural network.

With the development of Synthetic Aperture Radar (SAR) technology, automatic target recognition (ATR) is becoming increasingly important. In this paper, we proposed a 3-class target classification system in SAR images. The system is based on invariant wavelet moments and support vector machine (SVM) algorithm. It is a two-stage approach. The first stage is to extract and select a small set of wavelet invariant moment features to indicate target images. The wavelet invariant moments take both advantages of the wavelet inherent property of multi-resolution analysis and moment invariants quality of invariant to translation, scaling changes and rotation. The second stage is classification of targets with SVM algorithm. SVM is based on the principle of structural risk minimization (SRM), which has been shown better than the principle of empirical risk minimization (ERM) which is used by many conventional networks. To test the performance and efficiency of the proposed method, we performed experiments on invariant wavelet moments, different kernel functions, 2-class identification, and 3-class identification. Test results show that wavelet invariant moments indicate the target effectively; linear kernel function achieves better results than other kernel functions, and SVM classification approach performs better than conventional nearest distance approach.

In this paper, a novel color space transform is presented. It is an adaptive transform based on the application of independent component analysis to the RGB data of an entire color image. The result is a linear and reversible color space transform that provides three new coordinate axes where the projected data is as much as statistically independent as possible, and therefore highly uncorrelated. Compared to many non-linear color space transforms such as the HSV or CIE-Lab, the proposed one has the advantage of being a linear transform from the RGB color space, much like the XYZ or YIQ. However, its adaptiveness has the drawback of needing an estimate of the transform matrix for each image, which is sometimes computationally expensive for larger images due to the common iterative nature of the independent component analysis implementations. Then, an image subsampling method is also proposed to enhance the novel color space transform speed, efficiency and robustness. The new color space is used for a large set of test color images, and it is compared to traditional color space transforms, where we can clearly visualize its vast potential as a promising tool for segmentation purposes for example.

This paper describes a neural network based multiscale image restoration approach. Multilayer perceptrons are trained with artificial images of degraded gray level circles, in an attempt to make the neural network learn inherent space relations of the degraded pixels. The present approach simulates the degradation by a low pass Gaussian filter blurring operation and the addition of noise to the pixels at pre-established rates. The training process considers the degraded image as input and the non-degraded image as output for the supervised learning process. The neural network thus performs an inverse operation by recovering a quasi non-degraded image in terms of least squared. The main difference of the approach to existing ones relies on the fact that the space relations are taken from different scales, thus providing relational space data to the neural network. The approach is an attempt to come up with a simple method that leads to an optimum solution to the problem. Considering different window sizes around a pixel simulates the multiscale operation. In the generalization phase the neural network is exposed to indoor, outdoor, and satellite degraded images following the same steps use for the artificial circle image.

Template matching in real-time is a fundamental issue in many applications in computer vision such as tracking, stereo vision and autonomous navigation. The goal of this paper is present a system for automatic landmarks recognition in video frames over a georeferenced high resolution satellite image, for autonomous aerial navigation research. The video frames employed were obtained from a camera fixed to a helicopter in a low level flight, simulating the vision system of an unmanned aerial vehicle (UAV). The landmarks descriptors used in recognition task were texture features extracted by a Gabor Wavelet filters bank. The recognition system consists on a supervised neural network trained to recognize the satellite image landmarks texture features. In activation phase, each video frame has its texture feature extracted and the neural network has to classify it as a predefined landmark. The video frames are also preprocessed to reduce their difference of scale and rotation from the satellite image before the texture feature extraction, so the UAV altitude and heading for each frame are considered as known. The neural network techniques present the advantage of low computational cost, been appropriate to real-time applications. Promising results were obtained, mainly during flight over urban areas.

Gesture recognition is a new human-machine interface method implemented by pattern recognition(PR).In order to assure robot safety when gesture is used in robot control, it is required to implement the interface reliably and accurately. Similar with other PR applications, 1) feature selection (or model establishment) and 2) training from samples, affect the performance of gesture recognition largely. For 1), a simple model with 6 feature points at shoulders, elbows, and hands, is established. The gestures to be recognized are restricted to still arm gestures, and the movement of arms is not considered. These restrictions are to reduce the misrecognition, but are not so unreasonable. For 2), a new biological network method, called lobe component analysis(LCA), is used in unsupervised learning. Lobe components, corresponding to high-concentrations in probability of the neuronal input, are orientation selective cells follow Hebbian rule and lateral inhibition. Due to the advantage of LCA method for balanced learning between global and local features, large amount of samples can be used in learning efficiently.

As reported in our publications in SPIE conferences in the last two years, we can use a simple VB6 program to break the boundaries of some selected objects in an edge-detected binary picture into many simple branches and reconstruct accurately the boundaries of the original objects free of noise. By this means we can then program the computer to automatically learn some standard objects and automatically recognize any test objects by a novel topological connections of the simple, bi-directional graph of the object boundaries. It is very accurate, yet very robust, way to recognize the test objects because it is like the design of an electric circuit. When the way of connection of an electric circuit, or the topology of an electric circuit, is fixed, all the electrical properties of the circuit are fixed.

Dot diffusion is a halftoning technique that is based on the traditional error diffusion concept, but offers a high degree of parallel processing by its block based approach. Traditional dot diffusion however suffers from periodicity artifacts. To limit the visibility of these artifacts, we propose grid diffusion, which applies different class matrices for different blocks. Furthermore, in this paper we will discuss two approaches in the dot diffusion framework to generate green-noise halftone patterns. The first approach is based on output dependent feedback (hysteresis), analogous to the standard green-noise error diffusion techniques. We observe that the resulting halftones are rather coarse and highly dependent on the used dot diffusion class matrices. In the second approach we don't limit the diffusion to the nearest neighbors. This leads to less coarse halftones, compared to the first approach. The drawback is that it can only cope with rather limited cluster sizes. We can reduce these drawbacks by combining the two approaches.

This paper presents parameter estimation techniques useful for detecting background changes in a video sequence with extreme foreground activity. A specific application of interest is automated detection of the covert placement of threats (e.g., a briefcase bomb) inside crowded public facilities. We propose that a histogram of pixel intensity acquired from a fixed mounted camera over time for a series of images will be a mixture of two Gaussian functions: the foreground probability distribution function and background probability distribution function. We will use Pearson's Method of Moments to separate the two probability distribution functions. The background function can then be "remembered" and changes in the background can be detected. Subsequent comparisons of background estimates are used to detect changes. Changes are flagged to alert security forces to the presence and location of potential threats. Results are presented that indicate the significant potential for robust parameter estimation techniques as applied to video surveillance.

This study proposes a new approach to perform motion estimation on a set of feature points via elastic graph matching. The approach starts by constructing a labeled graph on a set of feature points in the first image of a given sequence, and then continues to sequentially match the graph with the remaining images in this sequence. The matching is based on a similarity function that depends on image brightness and motion characteristics on one side, and on geometric distortion on the other side. The main advantage of the proposed approach is that it preserves high-level image characteristics outlined by the geometrical structure of moving objects and their relative positions in space and time, while it simultaneously accounts for both low-level measurements (motion and intensity).

The resizing of data, either upscaling or downscaling based on need for increased or decreased resolution, is an important signal processing technique due to the variety of data sources and formats used in today's world. Image interpolation, the 2D variation, is commonly achieved through one of three techniques: nearest neighbor, bilinear interpolation, or bicubic interpolation. Each method comes with advantages and disadvantages and selection of the appropriate one is dependent on output and situation specifications. Presented in this paper are algorithms for the resizing of images based on the analysis of the sum of primary implicants representation of image data, as generated by a logical transform. The most basic algorithm emulates the nearest neighbor technique, while subsequent variations build on this to provide more accuracy and output comparable to the other traditional methods. Computer simulations demonstrate the effectiveness of these algorithms on binary and grayscale images.

This paper presents a fuzzy system approach using texture and color to classify living coral cover in underwater color images acquired by an autonomous underwater vehicle (AUV). The proposed fuzzy system for classification consists in the assigning of fuzzy memberships to different image features such as the mean, the spatial variance, the Gabor filter response standard deviation, and the wavelet energy. These fuzzy sensors are applied to the different segments present in the images. The segmentation of the images is previously done using the Homogeneity Coefficient Segmentation Algorithm (LHC). The resulted classification of the regions is compared against ground truth maps of the images. A correct classification over 80% was achieved in two different 25 images sets of two different areas.

In different applications, it is often desirable to retrieve useful information from multichannel (color, multispectral, dual or full-polarization) images. On one hand, multichannel images are potentially able to provide a lot of useful information about sensed objects (terrains). On the other hand, the task of its reliable extraction is very complicated. And there are many reasons behind this like inherent noise, lack of a priori information about object features, complexity of scenes, etc. Therefore, numerous different approaches based on various functional principles and mathematical background have been already put forward. In majority of them, image classification and segmentation are common operations that precede estimation of object parameters. However, practically all methods are far away from completeness and/or perfection since they suffer from different drawbacks and application restrictions. Recently we have proposed methods based on learning with local parameter clustering that were rather successfully applied to image locally adaptive filtering and detection of objects with certain properties. This paper is an attempt to extend this approach to image classification, segmentation and object parameter estimation. A particular application of substance quantitative analysis from color images is considered. The proposed approach is shown to solve the aforementioned task quite well and to have a rather high potential for other applications.

Time varying motion blur is sometimes brought by camera shake, and alternate blurred frames and sharp frames are included in video clips. This degradation is considered a kind of breathing distortion, which means oscillatory changes of blur. We propose a motion deblurring method which aims to suppress breathing distortions. A spatio-temporal regularization is introduced into the deconvolution approach to smooth the temporal change of blur.

In this paper we present the capability of the Rank M-Type Radial Basis Function (RMRBF) Neural Network in medical image processing applications. The proposed neural network uses the proposed RM-estimators in the scheme of radial basis function to train the neural network. The RMRBF-based training is less biased by the presence of outliers in the training set and was proved an accurate estimation of the implied probabilities. Other RBF based algorithms were compared with our approach in pdf estimation on the microcalcification detection in mammographic image analysis. From simulation results we observe that the RMRBF gives better estimation of the implied pdfs and has show better classification capabilities.

Vegetables are widely planted all over China, but they often suffer from the some diseases. A method of major technical and economical importance is introduced in this paper, which explores the feasibility of implementing fast and reliable automatic identification of vegetable diseases and their infection grades from color and morphological features of leaves. Firstly, leaves are plucked from clustered plant and pictures of the leaves are taken with a CCD digital color camera. Secondly, color and morphological characteristics are obtained by standard image processing techniques, for examples, Otsu thresholding method segments the region of interest, image opening following closing algorithm removes noise, Principal Components Analysis reduces the dimension of the original features. Then, a recently proposed boosting algorithm AdaBoost. M2 is applied to RBF networks for diseases classification based on the above features, where the kernel function of RBF networks is Gaussian form with argument taking Euclidean distance of the input vector from a center. Our experiment performs on the database collected by Chinese Academy of Agricultural Sciences, and result shows that Boosting RBF Networks classifies the 230 cucumber leaves into 2 different diseases (downy-mildew and angular-leaf-spot), and identifies the infection grades of each disease according to the infection degrees.