In this paper a general architecture for the computer-aided reconstruction of strip-cut shredded documents is presented. The matching of the remnants is performed on the base of the visual content of the strips, described by means of automatically extracted numerical features. A clustering approach is adopted in order to reduce progressively the dimension of the sets of remnants in which the exhaustive search for the matching need to be performed.

This paper presents the results of comparing two digital images acquired using two different light sources. One of the sources is a 50-W metal halide lamp located in the compartment of an industrial borescope and the other is a 1 W LED placed at the tip of the insertion tube of the borescope. The two images are compared quantitatively and qualitatively using feature extraction and luminance matching approaches. Quantitative methods included the images' histograms, intensity profiles along a line segment, edges, and luminance measurement. Qualitative methods included image registration and linear conformal transformation with eight control points. This transformation is useful when shapes in the input image are unchanged, but the image is distorted by some combination of translation, rotation, and scaling. The gray-level histogram, edge detection, image profile and image registration do not offer conclusive results. The LED light source, however, produces good images for visual inspection by the operator. The paper presents the results and discusses the usefulness and shortcomings of various comparison methods.

While analyzing a scene of interest in real environments, the acquisition and evaluation of image series has proven to yield promising results in order to provide useful information. However, acquiring and evaluating image series imposes several difficulties on the imaging and analysis process: The amount of data to be processed increases significantly, especially when more than one parameter is varied. Recording image series thus leads to a dilemma: Whereas a dense scanning of a varied parameter is desirable in order not to lose any information of interest, the number of recorded images should be as small as possible to ensure both adequate acquisition time and manageable amount of data. This dilemma can be considered as a sampling issue of the parameter space of a variable image acquisition. However, when two or more parameters are varied simultaneously, the resulting sampling condition is not necessarily a simple superposition of the one-dimensional cases. In this contribution, the topic of optimally sampling the parameter spaces for image series is addressed. Sampling conditions for different parameters to be varied are discussed. For image series with multiple varied parameters, interdependences that can be used to reduce the acquisition expenses without loosing relevant information are pointed out.

A novel solution for automatic hardwood inspection is presented. A sophisticated multi sensor system is required for reliable results. Our system works on a data stream of more than 50 MByte/Sec in input and up to 100 MByte/Sec inside the processing queue. The algorithm is divided into multiple steps. Along a fixed grid the images are decomposed into small squares. 55 texture- and color features are computed for each square. A Maximum Likelihood classifier assigns each square to one out of 12 defect classes with a recognition rate better than 97%. Depending on the defect type a dedicated threshold operation is performed for segmentation. Threshold levels and the selection of the input channel (RGB + filtered images) is the result of the former classification step. A fast algorithm computes bounding rectangles from blobs. Defect type dependent rules are used to combine rectangles. Two additional fast high resolution 3D measurement systems add board shape and 3D defect information. All defect rectangles are passing an additional plausibility check in the last data fusion process before they are delivered to the optimization computer. To guarantee a short response time, image acquisition and image processing are performed in parallel on parallel computing hardware.

Two-photon microscopy in combination with novel fluorescent labeling techniques enables imaging of three-dimensional neuronal morphologies in intact brain tissue. In principle it is now possible to automatically reconstruct the dendritic branching patterns of neurons from 3D fluorescence image stacks. In practice however, the signal-to-noise ratio can be low, in particular in the case of thin dendrites or axons imaged relatively deep in the tissue. Here we present a nonlinear anisotropic diffusion filter that enhances the signal-to-noise ratio while preserving the original dimensions of the structural elements. The key idea is to use structural information in the raw data -- the local moments of inertia -- to locally control the strength and direction of diffusion filtering. A cylindrical dendrite, for example, is effectively smoothed only parallel to its longitudinal axis, not perpendicular to it. This is demonstrated for artificial data as well as for in vivo 2-photon microscopic data from pyramidal neurons of rat neocortex. In both cases noise is averaged out along the dendrites, leading to bridging of apparent gaps, while dendritic diameters are not affected. The filter is a valuable general tool for smoothing cellular processes and is well suited for preparing data for subsequent image segmentation and neuron reconstruction.

In this paper, we first analyze a computationally efficient nonlinear edge-preserving smoothing filter, the sigma filter, in the framework of robust estimation. The sigma filter is nearly W-estimator using the metric trimming influence function with a subtle difference. Based on the analysis, we further develop another form of sigma filter as M-estimator of robust estimation using the same metric trimming influence function. The new form of sigma filter is more suitable for hardware implementation. We also compare the sigma filter with other nonlinear edge-preserving filters such as bilateral filter and median filter, in the framework of robust estimation. Finally, experimental results are reported and discussed.

Air Force Research Labs, Sensors Directorate has constructed and tested a coherent imaging system. The Laservision coherent imaging system resolution goals are compatible with long-range target identification based on the image characteristics of the target. The system received reflected coherent light from the laser using an optical telescope, which fed a CCD detector to collect the scene intensity. Registration of individual images remains a significant problem in the generation of accurate images collected using coherent imaging systems. An investigation of the performance of an image registration algorithm was conducted using data collected from a coherent optical imaging. The algorithm under study was implemented on a general-purpose computer running the MATLAB simulation environment. This paper documents the performance of the proposed technique compared to that of the cross-correlation algorithm.

Image filtering or denoising is a problem widely addressed in optical, infrared and radar remote sensing data processing. Although a large number of methods for image denoising exist, the choice of a proper, efficient filter is still a difficult problem and requires wide a priori knowledge. Locally adaptive filtering of images is an approach that has been widely investigated and exploited during recent 15 years. It has demonstrated a great potential. However, there are still some problems in design of locally adaptive filters that is generally too heuristic. This paper puts forward a new approach to get around this shortcoming. It deals with using learning with clustering in order to make the procedure of locally adaptive filter design more automatic and less subjective. The performance of this approach to learning and locally adaptive filtering has been tested for mixed Gaussian multiplicative+impulse noise environment. Its advantages in comparison to another learning methods and the efficiency of the considered component filters is demonstrated by both numerical simulation data and real-life radar image processing examples.

Shape and texture features have been used for some time for pattern recognition in datasets such as remote sensed imagery, medical imagery, photographs, etc. In this paper, we investigate shape and texture features for pattern recognition in simulation data. In particular, we explore which features are suitable for characterizing regions of interest in images resulting from fluid mixing simulations. Three texture features -- gray level co-occurrence matrices, wavelets, and Gabor filters -- and two shape features -- geometric moments and the angular radial transform -- are compared. The features are evaluated using a similarity retrieval framework. Our preliminary results indicate that Gabor filters perform the best among the texture features and the angular radial transform performs the best among the shape features. The feature which performs the best overall is dependent on how the groundtruth dataset is created.

A novel Markov random field (MRF) based framework is developed for the problem of 3D object recognition in multiple statuses. This approach utilizes densely sampled grids to represent the local information of the input images. Markov random field models are then created to model the geometric distribution of the object key points. Flexible matching, which seeks to find an accurate correspondence mapping between the key points of two images, is performed by combining the local similarities with the geometric relations using the highest confidence first (HCF) method. Afterwards, similarities between different images are calculated for object recognition. The algorithm is evaluated using the Coil-100 object database. The excellent recognition rates achieved in all the experiments indicate that our approach is well-suited for appearance-based 3-D object recognition. Comparisons with previous methods show that the proposed one is far more robust in the presence of object zooming, rotation, occlusion, noise, and viewpoint variations.

Improvement of the accuracy of an automatic car license plate recognition system is studied. The recognition system has been installed in several ports, factories and national borders around Europe and Middle East. The paper describes the anatomy of a portable version of the recognition system, in which the requirements are different from a fixed installation. For example, the portable system recognizes the license plates with no external triggers (such as inductive loops) and the installation is left to the end-user. The placement of the recognition camera in such systems is often a compromise and cannot be fully controlled. The system learns the characteristics of the setup and tries to correct the imperfections that are due to the installation by a non-expert. One of the main problems is the angle in which the camera is looking at the road, which causes the plate and the characters appear skewed. This paper proposes an algorithm for cancelling the skewness effect. The skewness parameters can be estimated from the earlier recognition results, and the proposed system learns to correct the skewness resulting in improved recognition results. The improvement of the recognition accuracy is illustrated by experiments with an annotated database of license plate images.

The automatic detection of Region of Interests (ROI) is an active research area in the design of machine vision systems. By using bottom-up image processing algorithms to predict human eye fixations or focus of attention and extract the relevant embedded information content in images has been widely applied in this area, especially in mobile robot navigation. Text that appears in images contains large quantities of useful information. Further more, many potential landmarks in mobile robot navigation contain text, such as nameplates, information signs and hence scene text is an important feature to be extracted. In this paper, we propose a simple and fast text localization algorithm based on a zero-crossing operator, which can effectively detect text-based features in an indoor environment for mobile robot navigation. This method is based on the idea that high local spatial variance is one of the distinguishing characteristics of text. Text in images has distinct intensity/color differences relative to its neighbourhood background and appears in clusters with uniform inter-character distance. If we compute the spatial variance along the text line we can get a large value, while the spatial variance in the background is fairly low. Experimental results show that calculating the spatial variance to detect text-based landmarks in real-time is an effective and efficient method for mobile robot navigation.

The best basis paradigm is a lower cost alternative to the principal component analysis (PCA) for feature extraction in pattern recognition applications. Its main idea is to build a collection of bases and search for the best one in terms of e.g. best class separation. Recently, fast best basis search algorithms have been generalized for anisotropic wavelet packet bases. Anisotropy is preferable for 2-D objects since it helps capturing local image features in a better way. In this contribution, the best anisotropic basis search framework is applied to the problem of recognition of characters captured from gray-scale pictures of car license plates. The goals are to simplify the classifier and to avoid a preliminary binarization stage by extracting features directly from the gray-scale images. The collection of bases is formed by anisotropic wavelet packets. The search algorithm seeks for a basis providing the lowest-dimensional data representation preserving the inter-class separability for given training data set, measured as Euclidean distance between class centroids. The relationship between the feature extractor and classifier complexity is clarified by training neural networks for different local bases. The proposed methodology shows its superiority to PCA as it yields equal and even lower classification error rate with considerably reduced computational costs.

In this paper, we propose an efficient and fast feature extraction method for iris recognition using wavelet transforms. The wavelet transforms have good space-frequency localization and there exist a number of fast algorithms. In particular, the coefficients of the lowest frequency band, reflecting the characteristics of the whole iris pattern, are used as a feature vector. However, a major problem of iris recognition is that noise such as the eyelid, the eyebrow and glint may be included in iris texture. Such noises adversely affect the performance of iris recognition systems. In order to solve these problems, after dividing the iris texture into a number of sub-regions, we propose to apply the wavelet transform separately to each sub-region and to extract a feature vector from each sub-region. In matching module, we discard some sub-regions which have large differences to exclude a potential noise. Experiments were performed using 3136 eye images acquired from 94 individuals. Experimental results show that the performance of proposed method is comparable to that of the method using Gabor transform and region division noticeably improves recognition performance for both methods. However, it is noted that the processing time of the former is much faster than that of the latter.

This paper presents a method to predict the limit of possible resolution enhancement given a sequence of low-resolution images. Three important parameters influence the outcome of this limit: the total Point Spread Function (PSF), the Signal-to-Noise Ratio (SNR) and the number of input images. Although a large number of input images captured by a system with a narrow PSF and a high SNR are desirable, these conditions are often not achievable simultaneously. To improve the SNR, cameras are designed with near optimal quantum efficiency and maximum fill-factor. However, the latter widens the system PSF, which puts more weight on the deblurring part of a super-resolution (SR) reconstruction algorithm. This paper analyzes the contribution of each input parameters to the SR reconstruction and predicts the best attainable SR factor for given a camera setting. The predicted SR factor agrees well with an edge sharpness measure computed from the reconstructed SR images. A sufficient number of randomly positioned input images to achieve this limit for a given scene can also be derived assuming Gaussian noise and registration errors.

We consider a signal restoration from observations corrupted by random noise. The local maximum likelihood technique allows to deal with quite general statistical models of signal dependent observations, relaxes the standard parametric modelling of the standard maximum likelihood, and results in flexible nonparametric regression estimation of the signal. We deal with the anisotropy of the signal using multi-window directional sectorial local polynomial approximation. The data-driven sizes of the sectorial windows, obtained by the intersection of confidence interval (ICI) algorithm, allow to form starshaped adaptive neighborhoods used for the pointwise estimation. The developed approach is quite general and is applicable for multivariable data. A fast adaptive algorithm implementation is proposed. It is applied for photon-limited imaging with the Poisson distribution of data. Simulation experiments and comparison with some of the best results in the field demonstrate an advanced performance of the developed algorithms.

Approaches analyzing local characteristics of an image prevail in image restoration. However, they are less effective in cases of restoring images degraded by large size point spread functions (PSFs) and heavy noise. In this paper, we propose a set theoretic approach to object-based image restoration that involves the following issues: representing the common characteristics of a class of objects that images of interest contain, the formulation to combine prior knowledge of the object, and the algorithm to find the solution. The common characteristics of objects are represented as deterministic sets built on principal component analysis based models of objects. Combining these sets with those arising from observed data, object-based image restoration is formulated in a set theoretic framework. Finally, a parallel subgradient projection algorithm is applied to find the intersection of the sets. Experiments performed on frontal face images using the proposed approach show significant improvement in large size and heavy noise degradation, compared with traditional methods based on local analysis. The proposed approach opens the possibilities of introducing the more prior knowledge pertaining to objects in terms of deterministic sets and solving the problem by abundant numerical algorithms under set theoretic framework.

Quality control in mammographic facilities has to be done periodically in order to ensure that the mammographic chain works properly. In particular global image quality is evaluated from a mammographic phantom film. A phantom is an object with the same anatomic shape and radiological response as an average dense-fleshed breast in which are embedded structures that mimic clinically relevant features such as microcalcifications, masses and fibers. This evaluation is done by visual observation of a phantom film and a global score is given depending on the number of objects seen by several observers. This paper presents the main results of a feasibility study of breast phantom scoring using digital image processing. First breast phantom films were digitized. For each category of structures, subimages were extracted from the digitized phantom. A noise reduction method was used as a pre-processing step. A local contrast enhancement was then applied. At last image segmentation was done. Noise reduction and contrast enhancement steps were both based on a direct contrast modification technique. Segmentation step was adapted to each embedded object. Nine digitized phantom films were studied and results show that an evaluation of mammographic facilities could be done using digital image processing.

In the framework of clinical respiratory investigation, providing accurate modalities for morpho-functional analysis is essential for diagnosis improvement, surgical planning and follow-up. This paper focuses on the upper airways investigation and develops an automated approach for 3D mesh reconstruction from MDCT acquisitions. In order to overcome the difficulties related to the complex morphology of the upper airways and to the image gray level heterogeneity of the airway lumens and thin cartilaginous septa, the proposed 3D reconstruction methodology combines 2D segmentation and 3D surface regularization approaches. The segmentation algorithm relies on mathematical morphology theory and provides robust discrimination of the airway lumen from the surrounding tissues, while preserving the connectivity relationship between the different anatomical structures. The 3D regularization step uses an energy-based modeling in order to achieve a smooth and well-fitted 3D surface of the upper airways. An accurate 3D mesh representation of the reconstructed airways makes it possible to develop specific clinical applications such as virtual endoscopy, surgical planning and computer assisted intervention. In addition, building up patient-specific 3D models of upper airways is highly valuable for the study and design of inhaled medication delivery via computational fluid dynamics (CFD) simulations.

New adaptive edge detection algorithms based on volumetric neighborhood size estimation for automatic three or higher dimensional biomedical image analysis are presented in this work. The proposed methods are based on nonparametric three-dimensional kernel functions obtained using the "three-term" orthogonal-type polynomial equations for different types of orthogonal polynomial families. The obtained multidimensional kernels can be of any volumetric neighborhood size and order of approximation. The optimal sizes of volume estimates, produced by the multidimensional convolution of the kernels with the multidimensional biomedical images, are controlled by a switch type variance dependent volume size selector. The proposed methods show excellent results in approximating the true position and shape of the edges of different organs of the human body represented in multidimensional biomedical images, which can have nonuniform voxel size and anisotropic image intensity and noise distribution.

The integer-to-integer discrete cosine and other unitary transforms become popular in recent years in such applications as lossless image coding, mobile computing, filter banks, and other areas. In this paper, we present new matrix representations of the reversible integer discrete cosine transforms (IDCT) that are based on the canonical representation and floor function. A new concept of the kernel integer discrete cosine transform is introduced, that allows us to reduce the calculation of the IDCT of type II to the kernel IDCT with a fewer operations of multiplication and floor function. The application of the kernel IDCT is described for calculation of the eight-point IDCT of type II, when seven multiplications and seven floor functions can be saved. The parameterized two-point DCT of type IV and its particular case that requires two operations of multiplication, four additions, and two floor functions are presented. The golden two-point DCT that minimizes the error of the cosine transform approximation by the IDCT is also considered. Application of the kernel DCT for calculating the eight-point IDCT results in the saving of twelve multiplications and twelve floor functions, when considered the decomposition of the transform by the Walsh-Hadamard transform.

An approach for estimating the distribution of soot particle dimensions from electron microscope images is studied. We have implemented simple image-analytical methods that produce an equivalent diameter distribution which can be compared with the corresponding distribution acquired via physical measurements. In comparison with manual object detection with conventional image processing software our method is time-saving and efficient. The shape of the particles emitted from the motor under different loads is affected by phenomena in exhaust dilution or release to air. Particle shape has a significant effect on its harmfulness to health. The researchers are also interested in knowing the actual particle size distribution to be able to improve catalyzer functionality. Engine exhaust particle emissions are often analyzed by methods based on the physical properties of soot particles, and assumptions about their size and shape. Our method provides data for refining these results. The implemented graphical user interface is semi-automatic and allows the user to remove erroneous results from the resulting thresholded image before the analysis. Then the task is to calculate the properties of interest over the particle population. We have written a toolbox with simple functions that realize the semi-automated analysis and the user interface for easy operation.

Adaptive steganographic techniques have become a standard direction taken when striving to complicate the detection of secret communication. The consideration of cover image features when embedding information is an effort to insert digital media while keeping the visual and the statistical properties of the cover image intact. There are several such embedding methods in existence today, applicable for different formats of images and with contrasting approaches. In this paper, we propose a new adaptive embedding technique which alters the least significant bit layers of an image. This technique is driven by three separate functions: (1) Adaptive selection of locations to embed. (2) Adaptive selection of number of bits per pixel to embed. (3) Adaptive selection of manner in which the information is inserted. Through the application of sensitive median-based statistical estimation and a recorded account of actions taken, the proposed algorithms are able to provide the desired level of security, both visually and statistically. In comparison with other methods offering the same level of security, the new technique is able to offer a greater embedding capacity. In addition, for the sake of thorough investigation and fair comparison, we will introduce a new stego capacity measure which will offer a means of comparing steganography methods applied across different formats of images. Finally, this new algorithm is created with the intention of implementing a new visual communication system acknowledging low energy consumption and limited computationally related resources.

Although the resolution of digital imaging device is rapidly increasing these days, digital photography still suffers from resolution limits when we want to enlarge a small picture. It is widely believed that edges play the important role when we evaluate the perceptual quality of resized images, and several edge-directed approaches have been proposed to soothe unwanted edge effects: jagging and blurring Edge-only-directed interpolations generally produce crispier edges than traditional interpolations such as bi-linear or bi-cubic, but they usually take longer time and sometimes create unacceptable defects where their edge estimation is incorrect. In this paper, we will address a new approach to compensate for blurring and aliasing effects using edge and corner cues in a given image.

In this paper we present a new online learning and classification algorithm and suggest its use for image segmentation. Our learning algorithm follows a variation of Bayesian estimation procedure, which combines prior knowledge and knowledge learned from data. Our classification algorithm strictly follows a statistical classification procedure. The new online learning algorithm is simple to implement, robust to initial parameters and has a linear complexity. The experimental results using computer generated data show that the proposed online learning algorithm can quickly learn the underlying structure from data. The proposed online learning algorithm is used to develop a novel image segmentation procedure. This image segmentation procedure is based on the region growing and merging approach. First, region growing is carried out using the online learning algorithm. Then, a merging operation is performed to merge the small regions. Two merging methods are proposed. The first method is based on statistical similarity and merges the statistically similar and spatially adjacent regions. The second method uses an information-based approach merging small regions into their neighbouring larger regions. Many experimental results clearly show the efficacy of the proposed image segmentation method.

In a recent paper, we have introduced a topological descriptor for shape representation based on classical Morse theory. More precisely, given a manifold M and a Morse function f on M, we build an invariant [Morse Shape Descriptor (MSD)] of the manifold from the ranks of relative homology groups of all pairs of lower levels of the function f. While the MSD is a robust invariant with very nice properties, its application requires time consuming computations of homology groups. We present a new and computationally efficient method to capture the essential of the information given by the MSD.

Object shape and camera motion can be recovered from a sequence of images using a set of feature point correspondences. This is known as the structure from motion problem. This paper describes a method of employing geometrical features available in a scene, in the form of straight lines, in a factorization-based structure from motion application. The effects of inaccuracies of feature data can be reduced by constraining the reconstructed features corresponding to the points forming straight lines. Our main contribution in this paper is the use of such geometric features to refine the shape recovery using the current advancements in the factorization method. Reconstructed features are mapped to straight lines and the measurement matrix containing image feature data is updated with the adjusted data. This increases the accuracy of reconstruction perceptually as well as quantitatively. The algorithm consists of first obtaining the reconstruction using singular value decomposition. Mapping 3D lines to sets of feature points is then carried out. The measurement matrix is refined followed by a second phase of factorization and, optionally, normalization to obtain metric reconstruction. Results pertaining to both synthetic and actual sequences of images are presented.

In this paper, a new strategy is presented to map the pixel space from infrared (IR) image to the pixel space of optical images. This is done by utilizing the usual technique of system identification using the input-output measurements and trying to fit a parametric model under a given fitness criterion. In this case, the IR image is taken as the input data and a similar image taken simultaneously by a normal CCD camera is taken as the output data. The acquired data are then fitted with an autoregressive moving average (ARMA) model in 1D under the H_∞ criterion which is the most robust and statistically less-dependent fitness criterion. The well known LMS algorithm can reach the H_∞ bounds under certain impositions. By structuring the given data to suite these impositions, the LMS algorithm is used to reach the optimal bound. This implies a faster convergence with better performance. The results have shown remarkable enhancement in the images and appears to be promising for the real-time applications.

Obtaining valid iris texture is the precondition of iris based personal verification and identification. With a special camera, illumination environment and users' cooperation play an important role in capturing high quality iris image. However, the captured iris images are usually corrupted by partial occlusion (by eyelids and eyelashes) and white spots (from specular refection). An efficient scheme of obtaining valid iris texture is described in this paper. The described scheme includes two parts: a new algorithm for locating iris texture and a novel algorithm for segmenting corrupted iris texture. In order to overcome the drawbacks of traditional iris texture location methods, which are sensitive to white spots and occlusion, a new location algorithm is proposed. The proposed location algorithm is able to locate iris texture rapidly by avoiding the tremendous computation demands imposed by Canny detector, Hough transformation and integrodifferential operator. In addition, a novel segmentation algorithm is proposed in the latter part of this paper to exclude the occlusion regions from the corrupted annular iris texture. Experiments show that the proposed scheme is efficient and robust for obtaining valid iris texture, even if partial occlusion and white spots appear. Moreover, the average time cost of our scheme is about 150ms on Pentium IV 2.8GHz PC, which satisfies the real-time requirement.

A new technique for motion blur removal is proposed in this paper. Motion blur occurs when there is relative motion between the camera and the object of interest. To remove the blur, it is first necessary to identify the point spread function (PSF). We study the characteristics of restoration errors due to some elaborately selected PSFs. The direction and extent of the PSF can be determined after some intentional restorations. Wiener filter is then incorporated to restore the blurred image. The algorithm is effective and robust to noise. We apply it to the restoration of motion blurred license plates images. The results are fairly satisfactory.

An image processing algorithm, mainly based on morphological enhancement and segmentation, is developed and applied to optical microscope images of transverse cuts of fabric filaments, to retrieve useful shape characteristics. Adaptive filtering and non-linear fitting algorithms are also applied. Computer generated noisy images are used to estimate the algorithm accuracy with excellent results. This algorithm is a significant improvement over the current human-based inspection method for filament shape analysis, and its development and application will improve quality control in textile industry. The complete procedure is outlined in the present work, showing relevant results and pointing out pertinent restrictions.

A region based illumination-normalization method which only uses gray level values in an image as information is proposed in this paper. The general purpose of illumination normalization is to reduce lighting effect when the testing images are captured in different environment and supplies useful and uniform data. We apply the algorithm in our face recognition system. The algorithm first divides the testing face images into two parts using homomorphic filtering method. One is the face feature, F₀, and the other one is the illumination information, I₀. Next, an illumination reference model is constructed from a set of normal face images. The illumination information of the testing face image is then adjusted according to the illumination reference model. The modified illumination information, I_α, and original face feature, F₀, are at last combined to make up the normalized face image. The face recognition result is improved after the algorithm is applied.

In this paper we propose a new approach for the synthesis of natural video textures. After generalizing the bidimensional extended self-similar (ESS) model to the three dimensional (3D) case we want to generate samples of 3D-ESS fields on a discrete grid. The video texture is modeled according to the 3D-ESS model. The autocorrelation functions (ACFs) of the increments of the original video texture, at different spatial and temporal scales, are estimated according to the 3D-ESS model. A synthetic 3D-ESS field, whose increments have the same ACFs of the corresponding ones of the given prototype, is generated using the incremental Fourier synthesis algorithm.

The task of object detection depends on the ability to suppress the noise present in images in order to increase the signal-to-noise ratio. The standard linear matched filter is the optimal filter on the assumption of the Gaussian distribution of the signal and the noise. However, as a rule the distribution of the signal in image processing is not Gaussian. The linear matched filter becomes sub-optimal. Any non-Gaussian distribution function can be closely approximated using the Gaussian Mixture Model (GMM). We use GMM to approximate the signal distribution function and derive the optimal filter by means of mean square error (MSE) minimization. The optimal non-linear filter is determined by the assumed signal distribution function. We use non-linear matched filtering for point source detection in astronomical images. We derive the GMM components by fitting the theoretical point source distribution function. The filtered images are subjected to image segmentation and subsequent point source detection. The non-linear matched filtering has been tested with simulated data and has been shown to significantly improve the quality of point source detection. Receiver operating characteristic technique has been used to evaluate performance of various Gaussian mixtures for point source detection. This algorithm is currently used for the Spitzer Spatial Telescope.

We propose a system that detects the current speaker in multi-speaker videoconferencing by using lip motion. First, the system detects the face and lip region of each of the candidate speakers using face color and shape information. Then, to detect the current speaker, it calculates the change between the current frame and the previous frame in lip region. To close-up the detected current speaker, we used two CCD cameras. One is a general CCD camera, the other is a PTZ camera controlled by RS-232C serial port. The experimental result is the proposed system capable of detecting the face of current speaker in a video feed with more than three people, regardless of orientation of the faces. With this system, it only takes 4 to 5 seconds to zoom in on the speaker from the initial reference image. Also, it is a more efficient image transmission system for such things as video conferencing and internet broadcasting because it offers a close up face image at a resolution of 320x240, while at the same time providing a whole background image.

In this contribution, we study the problem of lossless compression of Bayer pattern CFA data. Two main issues are addressed: how to pack (reorder) pixels from the red, green and blue color planes into a structure appropriate for the subsequent compression algorithm (structural transformation) and how to utilize possible correlations between colors. While structural transformation to the red and blue color planes is straightforward as they are directly downsampled onto compact rectangular grids, the quincunx sampling grid of the green pixels allows different separations. We explore three different methods for green pixels separation and compare their peculiarities in the light of the chosen prediction-based compression algorithm, i.e. JPEG-LS. Two color decorrelation approaches are proposed as well. In the first approach, simple difference between the red or blue pixel and the corresponding nearest green pixel is calculated. In the second approach, several nearest green pixels are used to estimate the real green pixel value for the particular red or blue pixel location. The performance of the proposed algorithm is tested for real CFA raw data and results in terms of compression ratio are presented.

Our research deals with a semi-automatic region-growing segmentation technique. This method only needs one seed inside the region of interest (ROI). We applied it for spinal cord segmentation but it also shows results for parotid glands or even tumors. Moreover, it seems to be a general segmentation method as it could be applied in other computer vision domains then medical imaging. We use both the thresholding simplicity and the spatial information. The gray-scale and spatial distances from the seed to all the other pixels are computed. By normalizing and subtracting to 1 we obtain the probability for a pixel to belong to the same region as the seed. We will explain the algorithm and show some preliminary results which are encouraging. Our method has low computational cost and very encouraging results in 2D. Future work will consist in a C implementation and a 3D generalisation.

The main difficulty in segmenting a cell image occurs when there are red blood cells touching the leukocyte. Similar brightness of the touched red blood cells with the leukocytes make the separation of the cytoplasm from the red blood cells quite difficult. Conventional approaches were based on the search of the concavities created by contact of two round boundaries as two points to be connected for the separation. Here, we exploit the fact that the boundary of the leukocytes normally has a round shape and a small portion of it is disconnected due to the touching red blood cells. Specifically, at an initial central point of the nucleus in the leukocyte, we can generate the largest possible circle that covers a circular portion of the composite of nucleus and cytoplasm areas. Then, by perturbing the initial central points and selecting only those central points that do not cross the boundary, we can cover most of interior regions in the nucleus and the cytoplasm, separating the leukocyte from the touching red blood cells.

In this paper two complex wavelet transforms, namely the Gabor wavelet transform and Kingsbury's Dual-Tree Complex wavelet transform (DT-CWT) are compared for their capabilities to extract facial features. The Gabor wavelets extract directional features from images and find frequent applications in computer vision problems of face detection and face recognition. The transform involves convolving an image with an ensemble of Gabor kernels, scale and directionally parameterized. As a result, a redundant image representation is obtained, where the number of transformed images is equal to the number of Gabor kernels used. However, repetitive convolution with 2-D Gabor kernels is a rather slow computational operation. The DT-CWT is a recently suggested transform, which provides good directional selectivity in six different fixed orientations at dyadic scales with the ability to distinguish positive and negative frequencies. It has a limited redundancy of four for images and is much faster than the Gabor transform to compute. Therefore, it arises as a good candidate to replace Gabor transform in applications, where the speed (i.e. on-line implementation) is a critical issue. We involve the two wavelet families in facial landmarks detection and compare their performance by statistical tests, e.g. by building Receiver Operating Characteristic (ROC) curves and by measuring the sensitivity of a particular feature extractor. We also compare results of Bayesian classification for the two families of feature extractors involved.

Addressing spots in microarray images and deriving expression values for corresponding genes are fundamental tasks in microarray image analysis. Reliable expression values can be obtained only if the spot locations are accurately known. Here, a novel approach for spot addressing in microarray images based on supervised learning is proposed. The aim is to locate each spot through classifying the image based on local features into spot centers and background using support vector machine classifier. The resulting spot location information is complemented through image processing methods in the post-processing phase. Our method, through searching locations for individual spots, enables accurate segmentation and extraction of expression values. The benefit of searching individual spots becomes clear in case of misaligned spots or spot rows.