This PDF file contains the front matter associated with SPIE-IS&T Proceedings Volume 6812, including the Title Page, Copyright information, Table of Contents, and the Conference Committee listing.

We virtually restore faded black and white photographic prints by the method of decomposing the image into a smooth component that contains edges and smoothed homogeneous regions, and a rough component that may include grain noise but also fine detail. The decomposition into smooth and rough components is achieved using a rational filter. Two approaches are considered; in one, the smooth component is histogram-stretched and then gamma corrected before being added back to a homomorphically filtered version of the rough component; in the other the image is initially gamma corrected and shifted towards white. Each approach presents improvements with respect to the previously separately explored techniques of gamma correction alone, and the stretching of the smooth component together with the homomorphical filtering of the rough component, alone. After characterizing the image with the help of the scatter plot of a 2D local statistic of the type (local intensity, local contrast), namely (local average, local standard deviation), the effects of gamma correction are studied as the effects on the scatter plot, on the assumption that the quality of the image is related to the distribution of data on the scatter plot. Also, the correlation coefficient between the local average and the local deviation on the one hand, and the global average of the image play important descriptor roles.

Multivariate images are now commonly produced in many applications. If their process is possible due to computers power and new programming languages, theoretical difficulties have still to be solved. Standard image analysis operators are defined for scalars rather than for vectors and their extension is not immediate. Several solutions exist but their pertinence is hardly linked to context. In the present paper we are going to get interested in segmentation of vector images also including a priori knowledge. The proposed strategy combines a decision procedure (where points are classified) and an automatic segmentation scheme (where regions are properly extracted). The classification is made using a Bayesian classifier. The segmentation is computed via a region growing method: the morphological Watershed transform. A direct computation of the Watershed transform on vector images is not possible since vector sets are not ordered. So, the Bayesian classifier is used for computing a scalar distance map where regions are enhanced or attenuated depending on their similitude to a reference shape: the current distance is the Mahalanobis distance. This combination allows to transfer the decision function from pixels to regions and to preserve the advantages of the original Watershed transform defined for scalar functions. The algorithm is applied for segmenting colour images (with a priori) and medical images, especially dermatology images where skin lesions have to be detected.

In this paper we propose a new technique to detect random-valued impulse noise in images. In this method, the noisy pixels are detected iteratively through several phases. In each phase, a pixel will be marked as a noisy pixel if it does not have sufficient number of similar pixels inside the neighborhood window. The size of the window increases over the phases, so does the sufficient similar neighbor criterion. After the detection phases, all noisy pixels will be corrected in a recovering process. We compare the performance of this method with other recently published methods, in terms of peak signal to noise ratio and perceptual quality of the restored images. From the simulation results we observe that this method outperforms all other methods at medium to high noise rates. The algorithm is very fast, providing consistent performance over a wide range of noise rates. It also preserves fine details of the image.

A common distortion in videos acquired in long range observation systems is image instability in form of chaotic local displacements of image frames caused by fluctuations in the refraction index of the atmosphere turbulence. At the same time, such videos, which are designed to present moving objects on a stable background, contain tremendous redundancy that potentially can be used for image stabilization and perfecting provided reliable separation of stable background from true moving objects. Recently, it was proposed to use this redundancy for resolution enhancement of turbulent video through elastic registration, with sub-pixel accuracy, of segments of video frames that represent stable scenes. This paper presents results of investigation, by means of computer simulation, into how parameters of such a resolution enhancement process affect its performance and its potentials and limitations.

A method for the detection of cracks in old paper photographs is presented. Cracks in photographic prints usually result from a folding of the paper support of the photograph; they often extend across the entire image, along a preferred orientation. A first clue we exploit for crack detection is the fact that old prints have a characteristic sepia hue, due to aging and to the type of processing used at the time; a break of the gelatin exposes the white support paper; likewise, a break of the paper causes a black region in the digitized image. Thus, cracks are usually achromatic; this fact can be used for their detection on a color space with an explicit hue component. A series of parallel microcracks that run along the direction of a main crack usually result as well; even though the gelatin may not be broken, the folds corresponding to these microcracks cause a set of image discontinuities, observable at a high-enough resolution. In an interactive process, the user indicates the ends of the crack on the frame of the photo and the algorithm detects the crack pixels. In addition to color, the algorithm uses a multidirectional, multiresolution Gabor approach and mathematical morphology. The resulting method provides crack detection with good performance, as evidenced by the corresponding Receiver Operating Characteristics (ROC) graphs.¹

We propose an image restoration technique exploiting regularized inversion and the recent block-matching and 3D filtering (BM3D) denoising filter. The BM3D employs a non-local modeling of images by collecting similar image patches in 3D arrays. The so-called collaborative filtering applied on such a 3D array is realized by transformdomain shrinkage. In this work, we propose an extension of the BM3D filter for colored noise, which we use in a two-step deblurring algorithm to improve the regularization after inversion in discrete Fourier domain. The first step of the algorithm is a regularized inversion using BM3D with collaborative hard-thresholding and the seconds step is a regularized Wiener inversion using BM3D with collaborative Wiener filtering. The experimental results show that the proposed technique is competitive with and in most cases outperforms the current best image restoration methods in terms of improvement in signal-to-noise ratio.

Automated Explosive Detection Systems (EDS) utilizing Computed Tomography (CT) generate a series of 2-D projections from a series of X-ray scans OF luggage under inspection. 3-D volumetric images can also be generated from the collected data set. Extensive data manipulation of the 2-D and 3-D image sets for detecting the presence of explosives is done automatically by EDS. The results are then forwarded to human screeners for final review. The final determination as to whether the luggage contains an explosive and needs to be searched manually is performed by trained TSA (Transportation Security Administration) screeners following an approved TSA protocol. The TSA protocol has the screeners visually inspect the resulting images and the renderings from the EDS to determine if the luggage is suspicious and consequently should be searched manually. Enhancing those projection images delivers a higher quality screening, reduces screening time and also reduces the amount of luggage that needs to be manually searched otherwise. This paper presents a novel edge detection algorithm that is geared towards, though not exclusive to, automated explosive detection systems. The goal of these enhancements is to provide a higher quality screening process while reducing the overall screening time and luggage search rates. Accurately determining the location of edge pixels within 2-D signals, often the first step in segmentation and recognition systems indicates the boundary between overlapping objects in a luggage. Most of the edge detection algorithms such as Canny, Prewitt, Roberts, Sobel, and Laplacian methods are based on the first and second derivatives/difference operators. These operators detect the discontinuities in the differences of pixels. These approaches are sensitive to the presence of noise and could produce false edges in noisy images. Including large scale filters, may avoid errors generated by noise, but often simultaneously eliminating the finer edge details as well. This paper proposes a novel pixels ratio based edge detection algorithm which is immune to noise. The new method compares ratios of pixels in multiple directions to an adaptive threshold to determine edges in different directions.

In this contribution, we propose an adaptive multiresolution denoising technique operating in the wavelet domain that selectively enhances object contours, extending a restoration scheme based on edge oriented wavelet representation by means of adaptive surround inhibition inspired by the human visual system characteristics. The use of the complex edge oriented wavelet representation is motivated by the fact that it is tuned to the most relevant visual image features. In this domain, an edge is represented by a complex number whose magnitude is proportional to its "strength" while phase equals the orientation angle. The complex edge wavelet is the first order dyadic Laguerre Gauss Circular Harmonic Wavelet, acting as a band limited gradient operator. The anisotropic sharpening function enhances or attenuates large/small edges more or less deeply, accounting for masking effects induced by textured background. Adapting sharpening to the local image content is realized by identifying the local statistics of natural and artificial textures like grass, foliage, water, composing the background. In the paper, the whole mathematical model is derived and its performances are validated on the basis of simulations on a wide data set.

The need for information extracted from remotely sensed data has increased in recent decades. To address this issue, research is being conducted to develop a complete multi-stage supervised object recognition system. The first stage of this system couples genetic programming with standard unsupervised clustering algorithms to search for the optimal preprocessing function. This manuscript addresses the quantification and the characterization of the uncertainty involved in the random creation of the first set of candidate solutions from which the algorithm begins. We used a Monte Carlo type simulation involving 800 independent realizations and then analyzed the distribution of the final results. Two independent convergence approaches were investigated: [1] convergence based solely on genetic operations (standard) and [2] convergence based on genetic operations with subsequent insertion of new genetic material (restarting). Results indicate that the introduction of new genetic material should be incorporated into the preprocessing framework to enhance convergence and to reduce variability.

The automatic detection of runway hazards from a moving platform under poor visibility conditions is a multifaceted problem. The general approach that we use relies on looking at several frames of the video imagery to determine the presence of objects. Since the platform is in motion during the acquisition of these frames, the first step in the process is to correct for platform motion. Extracting the scene structure from the frames is our next goal. To rectify, enhance the details and to remove fog we perform multiscale retinex followed by edge detection on the imagery. In this paper, we concentrate on the automatic determination of runway boundaries from the rectified, enhanced, and edge-detected imagery. We will examine the performance of edgedetection algorithms for images that have poor contrast, and quantify their efficacy as runway edge detectors. Additionally, we will define qualitative criteria to determine the best edge output image. Finally, we will find an optimizing parameter for the detector that would help us to automate the detection of objects on the runway and thus the whole process of hazard detection.

This paper presents some analytic tools for modeling and analyzing the performance of latent fingerprint matchers. While the tools and techniques presented are not necessarily new, the manner in which these are employed is believed to be novel. It is shown that using relatively simple techniques, valuable insights are provided into what is otherwise nearly an intractable problem. In addition, three other topics are touched upon: 1) We proved a thumb-nail sketch of NIST's ELFT project. The purpose of the ELFT project is to investigate the steps required to achieve a partial lights-out latent fingerprint processing capability; 2) Preliminary data obtained from this project are analyzed using the proposed techniques; 3) An analysis is provided for predicting the performance of a "fully-populated" system when measurements can realistically only be performed on a small subset of the full-up repository/background/gallery.

The quality of wood is of increasing importance in wood industry. One important quality aspect is the average annual ring width and its standard deviation that is related to the wood strength and stiffness. We present a camera based measurement system for annual ring measurements. The camera system is designed for outdoor use in forest harvesters. Several challenges arise, such as the quality of cutting process, camera positioning and the light variations. In the freshly cut surface of log end the annual rings are somewhat unclear due to small splinters and saw marks. In the harvester the optical axis of camera cannot be set orthogonally to the log end causing non-constant resolution of the image. The amount of natural light in forest varies from total winter darkness to midsummer brightness. In our approach the image is first geometrically transformed to orthogonal geometry. The annual ring width is measured with two-dimensional power spectra. The two-dimensional power spectra combined with the transformation provide a robust method for estimating the mean and the standard deviation of annual ring width. With laser lighting the variability due to natural lighting can be minimized.

Tomographic image reconstruction is computationally very demanding. In all cases the backprojection represents the performance bottleneck due to the high operational count and resulting high demand put on the memory subsystem. In this study, we present the implementation of a cone beam reconstruction algorithm on the Cell Broadband Engine (CBE) processor aimed at real-time applications. The cone-beam backprojection performance was assessed by backprojecting a half-circle scan of 512 projections of 1024² pixels into a volume of size 512³ voxels. The projections are acquired on a C-Arm scanner and directed in real time to a CBE-based platform for real-time reconstruction. The acquisition speed typically ranges between 17 and 35 projections per second. On a CBE processor clocked at 3.2 GHz, our implementation performs this task in ~13 seconds, allowing for real time reconstruction.

Adaptive filtering is a compute-intensive algorithm aimed at effectively reducing noise without blurring the structures contained in a set of digital images. In this study, we take a generalized approach for adaptive filtering based on seven oriented filters, each individual filter implemented by a two-dimensional (2D) convolution with a mask size of 11 by 11 pixels. Digital radiology workflow imposes severe real-time constraints that require the use of hardware acceleration such as provided by multicore processors. Implementing complex algorithms on heterogeneous multicore architectures is a complex task especially for taking advantage of the DMA engines. We have implemented the algorithm on a Cell Broadband Engine (CBE) processor clocked at 3.2 GHz using a generic framework for multicore processors. This implementation is capable of filtering images of 512² pixels at a throughput of 40 frames per second while allowing changing the parameters in real time. The resulting images are directed to the DR monitor or to the real-time computed tomography (CT) reconstruction engine.

Real-time knowledge of capsule volume of an organ provides a valuable clinical tool for 3D biopsy applications. It is challenging to estimate this capsule volume in real-time due to the presence of speckles, shadow artifacts, partial volume effect and patient motion during image scans, which are all inherent in medical ultrasound imaging. The volumetric ultrasound prostate images are sliced in a rotational manner every three degrees. The automated segmentation method employs a shape model, which is obtained from training data, to delineate the middle slices of volumetric prostate images. Then a "DDC" algorithm is applied to the rest of the images with the initial contour obtained. The volume of prostate is estimated with the segmentation results. Our database consists of 36 prostate volumes which are acquired using a Philips ultrasound machine using a Side-fire transrectal ultrasound (TRUS) probe. We compare our automated method with the semi-automated approach. The mean volumes using the semi-automated and complete automated techniques were 35.16 cc and 34.86 cc, with the error of 7.3% and 7.6% compared to the volume obtained by the human estimated boundary (ideal boundary), respectively. The overall system, which was developed using Microsoft Visual C++, is real-time and accurate.

We describe a set of methods to enable fully automated analysis of a novel label-free spinning-disc format microarray system. This microarray system operates in a dual-channel mode, simultaneously acquiring fluorescence as well as interferometric signals. The label-free interferometric component enables the design of robust gridding methods, which account for rotational effects difficult to estimate in traditional microarray image analysis. Printing of microarray features in a Cartesian grid is preferable for commercial systems because of the benefits of using existing DNA/protein printing technologies. The spinning disc operation of the microarray requires spatial transformation of Cartesian microarray features, from the reader/scanner frame of reference to the disc frame of reference. We describe a fast spatial transformation method with no measurable degradation in the quality of transformed data, for this purpose. The gridding method uses frequency-domain information to calculate grid spacing and grid rotation. An adaptive morphological segmentation method is used to segment microarray spots with variable sizes accurately. The entire process, from the generation of the raw data to the extraction of biologically relevant information, can be performed without any manual intervention, allowing for the deployment of high-throughput systems. These image analysis methods have enabled this microarray system to achieve superior sensitivity limits.

In this paper, we present an image understanding algorithm for automatically identifying and ranking different image regions into several levels of importance. Given a color image, specialized maps for classifying image content namely: weighted similarity, weighted homogeneity, image contrast and memory colors are generated and combined to provide a metric for perceptual importance classification. Further analysis yields a region ranking map which sorts the image content into different levels of significance. The algorithm was tested on a large database of color images that consists of the Berkeley segmentation dataset as well as many other internal images. Experimental results show that our technique matches human manual ranking with 90% efficiency. Applications of the proposed algorithm include image rendering, classification, indexing and retrieval.

We present a local area-based, discontinuity-preserving stereo matching algorithm that achieves high quality results near depth discontinuities as well as in homogeneous regions. To address the well-known challenge of defining appropriate support windows for local stereo methods, we use the anisotropic Local Polynomial Approximation (LPA) - Intersection of Confidence Intervals (ICI) technique. It can adaptively select a nearoptimal anisotropic local neighborhood for each pixel in the image. Leveraging this robust pixel-wise shape-adaptive support window, the proposed stereo method performs a novel matching cost aggregation step and an effective disparity refinement scheme entirely within a local high-confidence voting framework. Evaluation using the benchmark Middlebury stereo database shows that our method outperforms other local stereo methods, and it is even better than some algorithms using advanced but computationally complicated global optimization techniques.

Accurate and robust image change detection and motion segmentation has been of substantial interest in the image processing and computer vision communities. To date no single motion detection algorithm has been universally superior while biological vision systems are so adept at it. In this paper, we analyze image sequences using phase plots generated from sequential image frames and demonstrate that the changes in pixel amplitudes due to the motion of objects in an image sequence result in phase space behaviour resembling a chaotic signal. Recent research in neural signals have shown biological neural systems are highly responsive to chaos-like signals resulting from aperiodic forcing functions caused by external stimuli. We then hypothesize an alternative physics-based motion algorithm from the traditional optical flow algorithm. Rather than modeling the motion of objects in an image as a flow of grayscale values as in optical flow, we propose to model moving objects in an image scene as aperiodic forcing functions, impacting the imaging sensor, be it biological or silicon-based. We explore the applicability of some popular measures for detecting chaotic phenomena in the frame-wise phase plots generated from sequential image pairs and demonstrate their effectiveness on detecting motion while robustly ignoring illumination change.

Regardless the final targeted application (compression, watermarking, texture analysis, indexation, ...), image/video modelling in the DCT domain is generally approached by tests of concordance with some well known pdfs (like Gaussian, generalised Gaussian, Laplace, Rayleigh ...). Instead of forcing the images/videos to stick to such theoretical models, our study aims at estimating the true pdf characterising their behaviour. In this respect, we considered three intensively used ways of applying DCT, namely on whole frames, on 4x4 blocks, and on 8x8 blocks. In each case, we first prove that a law modelling the corresponding coefficients exists. Then, we estimate this law by Gaussian mixtures and finally we identify the generality of such model with respect to the data on which it was computed and to the estimation method it relies on.

Gradient operators are commonly used in edge detection. Usually, proper smoothing processing is performed on the original image when a gradient operator is applied. Generally, the smoothing processing is embedded in the gradient operator, such that each component of the gradient operator can be decomposed into some smoothing processing and a discrete derivative operator, which is defined as the difference of two adjacent values or the difference between the two values on the two sides of the position under check. When the image is smoothed, the edges of the main objects are also smoothed such that the differences of the adjacent pixels across edges are lowered down. In this paper, we define the derivative of f at a point x as f'(x)=g(x+Δx)-g(x-Δx), where g is the result of smoothing f with a smoothing filter, and Δx is an increment of x and it is properly selected to work with the filter. When Δx=2, sixteen gradient directions can be obtained and they provide a finer measurement than usual for gradient operators.

Natural compound eyes combine a small eye volume with a large field of view (FOV) at the cost of comparatively low spatial resolution. Based on these principles, an artificial apposition compound-eye imaging system has been developed. In this system the total FOV is given by the number of channels along one axis multiplied with the sampling angle between channels. In order to increase the image resolution for a fixed FOV the sampling angle is made small. However, depending on the size of the acceptance angle, the FOVs of adjacent channels overlap which causes a reduction of contrast in the overall image. In this work we study the feasibility of using digital post-processing methods for images obtained with a thin compound-eye camera to overcome this reduction in contrast. We chose the Wiener filter for the post-processing and carried out simulations and experimental measurements to verify its use.

The explosion of VoD and HDTV services opened a new direction in watermarking applications: compressed domain watermarking, promising at least tenfold speed increase. While sound technical approaches to this emerging field are already available in the literature, at our best knowledge the present paper is the first related theoretical study. It considers the ISO/IEC 14496-10:2005 standard (also known as MPEG-4 AVC) and objectively describes with information theory concepts (noisy channel, noise matrices) the effects of the real-life watermarking attacks (like rotations, linear and non-linear filtering, StirMark). All the results are obtained on a heterogeneous corpus of 7 video sequences summing up to about 3 hours.

Active site prediction, well-known for drug design and medical diagnosis, is a major step in the study and prediction of interactions between proteins. The specialized literature provides studies of common physicochemical and geometric properties shared by active sites. Among these properties, this paper focuses on the travel depth which takes a major part in the binding with other molecules. The travel depth of a point on the protein solvent excluded surface (SES) can be defined as the shortest path accessible for a solvent molecule between this point and the protein convex hull. Existing algorithms providing an estimation of this depth are based on the sampling of a bounding box volume surrounding the studied protein. These techniques make use of huge amounts of memory and processing time and result in estimations with precisions that strongly depend on the chosen sampling rate. The contribution of this paper is a surface-based algorithm that only takes samples of the protein SES into account instead of the whole volume. We show this technique allows a more accurate prediction, at least 50 times faster. A validation of this method is also proposed through experiments with a statistical classifier taking as inputs the travel depth and other physicochemical and geometric measures for active site prediction.

Using our recently modified curve-fitting-topological-coding (CFTC) computer program, we can automatically obtain a precise topological code to represent the topological property of a closely reconstructed boundary of a selected object in an edge-detected picture. This topological property is perhaps the most important property to be used for object identification. It is very accurate, yet very robust, because the topological property is independent of geometrical location, shape, size, orientation, and viewing angles. It is very accurate if two different objects to be differentiated or to be identified have different boundary topologies. Patch noise and obscuring noise can also be automatically eliminated as shown in some live experiments.

The iris is currently believed to be the most accurate biometric for human identification. The majority of fielded iris identification systems are based on the highly accurate wavelet-based Daugman algorithm. Another promising recognition algorithm by Ives et al uses Directional Energy features to create the iris template. Both algorithms use Hamming distance to compare a new template to a stored database. Hamming distance is an extremely fast computation, but weights all regions of the iris equally. Work from multiple authors has shown that different regions of the iris contain varying levels of discriminatory information. This research evaluates four post-processing similarity metrics for accuracy impacts on the Directional Energy and wavelets based algorithms. Each metric builds on the Hamming distance method in an attempt to use the template information in a more salient manner. A similarity metric extracted from the output stage of a feed-forward multi-layer perceptron artificial neural network demonstrated the most promise. Accuracy tables and ROC curves of tests performed on the publicly available Chinese Academy of Sciences Institute of Automation database show that the neural network based distance achieves greater accuracy than Hamming distance at every operating point, while adding less than one percent computational overhead.

This paper describes a new nonlinear joint fusion and anomaly detection technique for mine detection applications using two different types of sensor data (synthetic aperture radar (SAR) and Hyperspectral sensor (HS) data). A well-known anomaly detector so called the RX algorithm is first extended to perform fusion and detection simultaneously at the pixel level by appropriately concatenating the information from the two sensors. This approach is then extended to its nonlinear version. The nonlinear fusion-detection approach is based on the statistical kernel learning theory which explicitly exploits the higher order dependencies (nonlinear relationships) between the two sensor data through an appropriate kernel. Experimental results for detecting anomalies (mines) in hyperspectral imagery are presented for linear and nonlinear joint fusion and detection for a co-registered SAR and HS imagery. The result show that the nonlinear techniques outperform linear versions.

This paper presents a new image-interpolation approach where one can adjust edge sharpness and texture intensity according to one's taste. This approach is composed of the three stages. At the first stage, with the BV-G imagede-composition variational model, an image is represented as a product of its two components so that its separated structural component may correspond to a cartoon image-approximation and its separated texture components may collect almost all oscillatory variations representing textures, and the texture component can be amplified or attenuated according to user's taste. At the second stage, each separated component is interpolated with an interpolation method suitable to it. Since the structural component keeps sharp edges, its proper interpolation method is a TV-regularization super-resolution interpolation method that can restore frequency components higher than the Nyquist frequency and remove sample-hold blurs without producing ringing artifacts near edges. The texture component is an oscillatory function, and its proper interpolation method is a smoothness-regularization super-resolution interpolation method that can restore continuous variations and remove the blurs. At the final stage, the two interpolated components are combined. The approach enlarges images without not only blurring edges but also destroying textures, and removes blurs caused by the sample-hold and/or the optical low-pass filter without producing ringing artifacts.

Majority of image filtering techniques are designed under assumption that noise is of special, a priori known type and it is i.i.d., i.e. spatially uncorrelated. However, in many practical situations the latter assumption is not true due to several reasons. Moreover, spatial correlation properties of noise might be rather different and a priori unknown. Then the assumption that noise is i.i.d. under real conditions of spatially correlated noise commonly leads to considerable decrease of a used filter effectiveness in comparison to a case if this spatial correlation is taken into account. Our paper deals with two basic aspects. The first one is how to modify a denoising algorithm, in particular, a discrete cosine transform (DCT) based filter in order to incorporate a priori or preliminarily obtained knowledge of spatial correlation characteristics of noise. The second aspect is how to estimate spatial correlation characteristics of noise for a given image with appropriate accuracy and robustness under condition that there is some a priori information about, at least, noise type and statistics like variance (for additive noise case) or relative variance (for multiplicative noise). We also present simulation results showing the effectiveness (the benefit) of taking into consideration noise correlation properties.

Face recognition continues to meet significant challenges in reaching accurate results and still remains one of the activities where humans outperform technology. An attractive approach in improving face identification is provided by the fusion of multiple imaging sources such as visible and infrared images. Hyperspectral data, i.e. images collected over hundreds of narrow contiguous light spectrum intervals constitute a natural choice for expanding face recognition image fusion, especially since it may provide information beyond the normal visible range, thus exceeding the normal human sensing. In this paper we investigate the efficiency of hyperspectral face recognition through an in house experiment that collected data in over 120 bands within the visible and near infrared range. The imagery was produced using an off the shelf sensor in both indoors and outdoors with the subjects being photographed from various angles. Further processing included spectra collection and feature extraction. Human matching performance based on spectral properties is discussed.

Compressed video is the digital raw material provided by video-surveillance systems and used for archiving and indexing purposes. Multimedia standards have therefore a direct impact on such systems. If MPEG-2 used to be the coding standard, MPEG-4 (part 2) has now replaced it in most installations, and MPEG-4 AVC/H.264 solutions are now being released. Finely analysing the complex and rich MPEG-4 streams is a challenging issue addressed in that paper. The system we designed is based on five modules: low-resolution decoder, motion estimation generator, object motion filtering, low-resolution object segmentation, and cooperative decision. Our contributions refer to as the statistical analysis of the spatial distribution of the motion vectors, the computation of DCT-based confidence maps, the automatic motion activity detection in the compressed file and a rough indexation by dedicated descriptors. The robustness and accuracy of the system are evaluated on a large corpus (hundreds of hours of in-and outdoor videos with pedestrians and vehicles). The objective benchmarking of the performances is achieved with respect to five metrics allowing to estimate the error part due to each module and for different implementations. This evaluation establishes that our system analyses up to 200 frames (720x288) per second (2.66 GHz CPU).

DSA images suffer from challenges like system X-ray noise and artifacts due to patient movement. In this paper, we present a two-step strategy to improve DSA image quality. First, a hierarchical deformable registration algorithm is used to register the mask frame and the bolus frame before subtraction. Second, the resulted DSA image is further enhanced by background diffusion and nonlinear normalization for better visualization. Two major changes are made in the hierarchical deformable registration algorithm for DSA images: 1) B-Spline is used to represent the deformation field in order to produce the smooth deformation field; 2) two features are defined as the attribute vector for each point in the image, i.e., original image intensity and gradient. Also, for speeding up the 2D image registration, the hierarchical motion compensation algorithm is implemented by a multi-resolution framework. The proposed method has been evaluated on a database of 73 subjects by quantitatively measuring signal-to-noise (SNR) ratio. DSA embedded with proposed strategies demonstrates an improvement of 74.1% over conventional DSA in terms of SNR. Our system runs on Eigen's DSA workstation using C++ in Windows environment.

In this paper we propose several improvements to the original non-local means algorithm introduced by Buades et al. which obtains state-of-the-art denoising results. The strength of this algorithm is to exploit the repetitive character of the image in order to denoise the image unlike conventional denoising algorithms, which typically operate in a local neighbourhood. Due to the enormous amount of weight computations, the original algorithm has a high computational cost. An improvement of image quality towards the original algorithm is to ignore the contributions from dissimilar windows. Even though their weights are very small at first sight, the new estimated pixel value can be severely biased due to the many small contributions. This bad influence of dissimilar windows can be eliminated by setting their corresponding weights to zero. Using the preclassification based on the first three statistical moments, only contributions from similar neighborhoods are computed. To decide whether a window is similar or dissimilar, we will derive thresholds for images corrupted with additive white Gaussian noise. Our accelerated approach is further optimized by taking advantage of the symmetry in the weights, which roughly halves the computation time, and by using a lookup table to speed up the weight computations. Compared to the original algorithm, our proposed method produces images with increased PSNR and better visual performance in less computation time. Our proposed method even outperforms state-of-the-art wavelet denoising techniques in both visual quality and PSNR values for images containing a lot of repetitive structures such as textures: the denoised images are much sharper and contain less artifacts. The proposed optimizations can also be applied in other image processing tasks which employ the concept of repetitive structures such as intra-frame super-resolution or detection of digital image forgery.

Optical flow algorithms for estimating image local motion in video sequences are based on the first term Taylor series expansion approximation of image variations caused by motion, which requires computing image spatial derivatives. In this paper we report an analytical assessment of lower bounds of optical flow estimation errors defined by the accuracy of the Taylor series expansion approximation of image variations and results of experimental comparison of performance of known optical flow methods, in which image differentiation was implemented through different commonly used numerical differentiation methods and through DFT/DCT based algorithms for precise differentiation of sampled data. The comparison tests were carried out using simulated sequences as well as real-life image sequences commonly used for comparison of optical flow methods. The simulated sequences were generated using, as test images, pseudo-random images with uniform spectrum within a certain fraction of the image base band specified by the image sampling rate, the fraction being a parameter specifying frequency contents of test images. The experiments have shown that performance of the optical flow methods can be significantly improved compared to the commonly used numerical differentiation methods by using the DFT/ DCT-based differentiation algorithms especially for images with substantial high-frequency content.

Super-resolution (SR) video processing method with reduced block artifacts using conventional block-matching algorithm (BMA) is proposed. To get high-quality SR results, accurate motion vectors are necessary in registration process. For real applications, block-based motion estimators are widely used, which show block-based motion errors if their motion vectors are employed for the registration. Incorrectly registered pixels due to the block-based motion errors limit the image quality improvement of the SR processing and even degrade the results by causing block artifacts. To reduce the artifacts from the inaccurately registered pixels, a weighting function using three-dimensional confidence measure is proposed in this paper. The measure uses spatial and inter-channel analysis to suppress the weight on incorrectly registered pixels during the SR process. Motion-compensated pixel differences and motion vector variances between previous and current frames are utilized for spatial analysis, and motion vector variance with constant acceleration model and pixel difference variances through LR frames are used for inter-channel analysis. Experimental results show significantly improved results in error regions keeping enhanced quality with accurately registered pixels, when motion vectors are found by conventional BMAs.

This paper presents an efficient solution for digital images sharpening, the Adaptive Directional Sharpening with Overshoot Control (ADSOC), a method based on a high-pass filter able to perform a stronger sharpening in the detailed zones of the image, preserving the homogeneous regions. The basic objective of this approach is to reduce the undesired effects. The sharpening introduced along strong edges or into uniform regions could provide unpleased ringing artifacts and noise amplification, which are the most common drawbacks of the sharpening algorithms. The ADSOC allows to the user to choose the ringing intensity and it doesn't increase the isolated noisy pixel luminance value. Moreover, the ADSOC works the orthogonally respect to the direction of the edges in the blurred image, in order to yield a more effective contrast enhancement. The experiments showed good algorithm performances in terms of booth visual quality and computational complexity.

Due to the fast migration of high resolution displays for home and office environments there is a strong demand for high quality picture scaling. This is caused on the one hand by large picture sizes and on the other hand due to an enhanced visibility of picture artifacts on these displays [1]. There are many proposals for an enhanced spatial interpolation adaptively matched to picture contents like e.g. edges. The drawback of these approaches is the normally integer and often limited interpolation factor. In order to achieve rational factors there exist combinations of adaptive and non adaptive linear filters, but due to the non adaptive step the overall quality is notably limited. We present in this paper a content adaptive polyphase interpolation method which uses "offline" trained filter coefficients and an "online" linear filtering depending on a simple classification of the input situation. Furthermore we present a new approach to a content adaptive interpolation polynomial, which allows arbitrary polyphase interpolation factors at runtime and further improves the overall interpolation quality. The main goal of our new approach is to optimize interpolation quality by adapting higher order polynomials directly to the image content. In addition we derive filter constraints for enhanced picture quality. Furthermore we extend the classification based filtering to the temporal dimension in order to use it for an intermediate image interpolation.

Image scrambling is used to make images visually unrecognizable such that unauthorized users have difficulty decoding the scrambled image to access the original image. This article presents two new image scrambling algorithms based on Fibonacci p-code, a parametric sequence. The first algorithm works in spatial domain and the second in frequency domain (including JPEG domain). A parameter, p, is used as a security-key and has many possible choices to guarantee the high security of the scrambled images. The presented algorithms can be implemented for encoding/decoding both in full and partial image scrambling, and can be used in real-time applications, such as image data hiding and encryption. Examples of image scrambling are provided. Computer simulations are shown to demonstrate that the presented methods also have good performance in common image attacks such as cutting (data loss), compression and noise. The new scrambling methods can be implemented on grey level images and 3-color components in color images. A new Lucas p-code is also introduced. The scrambling images based on Fibonacci p-code are also compared to the scrambling results of classic Fibonacci number and Lucas p-code. This will demonstrate that the classical Fibonacci number is a special sequence of Fibonacci p-code and show the different scrambling results of Fibonacci p-code and Lucas p-code.

Prostate cancer is the most commonly diagnosed cancer in males in the United States and the second leading cause of cancer death. While the exact cause is still under investigation, researchers agree on certain risk factors like age, family history, dietary habits, lifestyle and race. It is also widely accepted that cancer distribution within the prostate is inhomogeneous, i.e. certain regions have a higher likelihood of developing cancer. In this regard extensive work has been done to study the distribution of cancer in order to perform biopsy more effectively. Recently a statistical cancer atlas of the prostate was demonstrated along with an optimal biopsy scheme achieving a high detection rate. In this paper we discuss the complete construction and application of such an atlas that can be used in a clinical setting to effectively target high cancer zones during biopsy. The method consists of integrating intensity statistics in the form of cancer probabilities at every voxel in the image with shape statistics of the prostate in order to quickly warp the atlas onto a subject ultrasound image. While the atlas surface can be registered to a pre-segmented subject prostate surface or instead used to perform segmentation of the capsule via optimization of shape parameters to segment the subject image, the strength of our approach lies in the fast mapping of cancer statistics onto the subject using shape statistics. The shape model was trained from over 38 expert segmented prostate surfaces and the atlas registration accuracy was found to be high suggesting the use of this method to perform biopsy in near real time situations with some optimization.

Enhancing an image in such a way that maintains image edges is a difficult problem. Many current methods for image enhancement either smooth edges on a small scale while improving contrast on a global scale or enhance edges on a large scale while amplifying noise on a small scale. One method which has been proposed for overcoming this is anisotropic diffusion, which views each image pixel as an energy sync which interacts with the surrounding pixels based upon the differences in pixel intensities and conductance values calculated from local edge estimates. In this paper, we propose a novel image enhancement method which makes use of these smoothed images produced by diffusion methods. The basic steps of this algorithm are: a) decompose an image into a smoothed image and a difference image, for example by using anisotropic diffusion or as in Lee's Algorithm [14]; b) apply two image enhancement algorithms, such as alpha rooting [7] or logarithmic transform shifting [15]; c) fuse these images together, for example by weighting the two enhanced images and summing them for the final image. Computer simulations comparing the results of the proposed method and current state-of-the-art enhancement methods will be presented. These simulations show the higher performance, both on the basis of subjective evaluation and objective measures, of the proposed method over current methods.

The Landsat-7 Enhanced Thematic Mapper Plus (ETM+) is the sensor payload on the Landsat-7 satellite imager (launched on April 15th, 1999) that is a derivative of the Landsat-4 and 5 Thematic Mapper (TM) land imager sensors. Scan Line Corrector (SLC) malfunctioning appeared onboard on May 31, 2003. The SLC-Off problem was caused by failure of the SLC which compensates for the forward motion of the satellite [1]. As ETM+ is still capable of acquiring images with the SLC-Off mode, the need of applying new techniques and using other data sources to reconstruct the missed data is a challenging for scientists and final users of remotely sensed images. One of the predicted future roles of the Advanced Land Imager (ALI) onboard the Earth Observer One (EO-1) is its ability to offer a potential technological direction for Landsat data continuity missions [2]. In this regard more than the purposes of the work as fabricating the gapped area in the ETM+ the attempt to evaluate the ALI imagery ability is another noticeable point in this work. In the literature there are several techniques and algorithms for gap filling. For instance local linear histogram matching [3], ordinary kriging, and standardized ordinary cokriging [4]. Here we used the Regression Based Data Combination (RBDC) in which it is generally supposed that two data sets (i.e. Landsat/ETM+ and EO-1/ALI) in the same spectral ranges (for instance band 3 ETM+ and band 4 ALI in 0.63 - 0.69 μm) will have meaningful and useable statistical characteristics. Using this relationship the gap area in ETM+ can be filled using EO-1/ALI data. Therefore the process is based on the knowledge of statistical structures of the images which is used to reconstruct the gapped areas. This paper presents and compares four regression based techniques. First two ordinary methods with no improvement in the statistical parameters were undertaken as Scene Based (SB) and Cluster Based (CB) followed by two statistically developed algorithms including Buffer Based (BB) and Weighted Buffer Based (WBB) techniques. All techniques are executed and evaluated over a study area in Sulawesi, Indonesia. The results indicate that the WBB and CB approaches have superiority over the SB and BB methods.

Prostate repeat biopsy has become one of the key requirements in today's prostate cancer detection. Urologists are interested in knowing previous 3-D biopsy locations during the current visit of the patient. Eigen has developed a system for performing 3-D Ultrasound image guided prostate biopsy. The repeat biopsy tool consists of three stages: (1) segmentation of the prostate capsules from previous and current ultrasound volumes; (2) registration of segmented surfaces using adaptive focus deformable model; (3) mapping of old biopsy sites onto new volume via thin-plate splines (TPS). The system critically depends on accurate 3-D segmentation of capsule volumes. In this paper, we study the effect of automated segmentation technique on the accuracy of 3-D ultrasound guided repeat biopsy. Our database consists of 38 prostate volumes of different patients which are acquired using Philips sidefire transrectal ultrasound (TRUS) probe. The prostate volumes were segmented in three ways: expert segmentation, semi-automated segmentation, and fully automated segmentation. New biopsy sites were identified in the new volumes from different segmentation methods, and we compared the mean squared distance between biopsy sites. It is demonstrated that the performance of our fully automated segmentation tool is comparable to that of semi-automated segmentation method.

We consider the problem of improving contour detection by filling gaps between collinear contour pieces. A fast algorithm is proposed which takes into account local edge orientation and local curvature. Each edge point is replaced by a curved elongated patch, whose orientation and curvature match the local edge orientation and edge. The proposed contour completion algorithm is integrated in a multiresolution framework for contour detection. Experimental results show the superiority of the proposed method to other well-established approaches.

In this paper a novel method for watermarking and ciphering color images is presented. The aim of the system is to allow the watermarking of encrypted data without requiring the knowledge of the original data. By using this method, it is also possible to cipher watermarked data without damaging the embedded signal. Furthermore, the extraction of the hidden information can be performed without deciphering the cover data and it is also possible to decipher watermarked data without removing the watermark. The transform domain adopted in this work is the Fibonacci-Haar wavelet transform. The experimental results show the effectiveness of the proposed scheme.

The deblurring of images corrupted by radial blur is studied. This type of blur appears in images acquired during an any camera translation having a substantial component orthogonal to the image plane. The point spread functions (PSF PSF) describing this blur are spatially varying. However, this blurring process does not mix together pixels lying on differen different radial lines, i.e. lines stemming from a unique point in the image, the so called "blur center". Thus, in suitable pola polar coordinates, the blurring process is essentially a 1-D linear operator, described by the multiplication with the blurrin blurring matrix. We consider images corrupted simultaneously by radial blur and noise. The proposed deblurring algorithm is base based on two separate forms of regularization of the blur inverse. First, in the polar domain, we invert the blurring matri matrix using the Tikhonov regularization. We then derive a particular modeling of the noise spectrum after both the regularize regularized inversion and the forward and backward coordinate transformations. Thanks to this model, we successfully use a denoisin denoising algorithm in the Cartesian domain. We use a non-linear spatially adaptive filter, the Pointwise Shape-Adaptive DCT, i in order to exploit the image structures and attenuate noise and artifacts. Experimental results demonstrate that the proposed algorithm can effectively restore radial blurred images corrupted by additive white Gaussian noise.

This paper proposes a novel data hiding scheme in which a payload is embedded into the discrete cosine transform domain. The characteristics of the Human Visual System (HVS) with respect to image fruition have been exploited to adapt the strength of the embedded data and integrated in the design of a digital image watermarking system. By using an HVS-inspired image quality metric, we study the relation between the amount of data that can be embedded and the resulting perceived quality. This study allows one to increase the robustness of the watermarked image without damaging the perceived quality, or, as alternative, to reduce the impairments produced by the watermarking process given a fixed embedding strength. Experimental results show the effectiveness and the robustness of the proposed solution.

Photometry of crowded fields is an old theme of astronomical image processing. Large space surveys in the UV (ultraviolet), like the GALEX mission (135-175 nm and 170-275 nm range), confronts us again with challenges like, very low light levels, poor resolution, variable stray-light in background, the extended and badly known PSFs (point spread functions), etc. However the morphological similitude of these UV images to their counterparts in the visible bands, suggests that we use all this high resolution data as the starting reference for the UV analysis. We choose the Bayesian approach. However there is not a straightforward way leading from the basic idea to its practical implementation. We will describe in this paper the path which starts with the original procedure (presented in a previous paper) and ends on the useful one. After a brief recall on the Bayesian method, we describe the process applied to restore from the UV images the point spread function (PSF) and the background due to stray-light. In the end we display the photometric performances reached for each channel and we discuss the consequences of the imperfect knowledge of background, the inaccuracy on object centring and on the PSF model. Results show a clear improvement by more than 2 mag on the magnitude limit and in the completeness of the measured objects relative to classical methods (it corresponds to more than 75000 new objects per GALEX field, i.e. approx 25% more). The simplicity of the Bayesian approach eased the analysis as well as the corrections needed in order to obtain a useful and reliable photometric procedure.

Face Recognition has been a major topic of research for many years and several approaches have been developed, among which the Principal Component Analysis (PCA) algorithm using Eigenfaces is the most popular. Eigenfaces optimally extract a reduced basis set that minimizes reconstruction error for the face class prototypes. The method is based on second-order pixel statistics and does not address higher-order statistical dependencies such as relationships among three or more pixels. Independent Component Analysis (ICA) is a recently developed linear transformation method for finding suitable representations of multivariate data, such that the components of the representation are as statistically independent as possible. The face image class prototypes in ICA are considered to be a linear-mixture of some unknown set of basis images that are assumed to be statistically independent, in the sense that the pixel values of one basis image cannot be predicted from that of another. This research evaluates the performance of ICA for face recognition under varying conditions like change of expression, change in illumination and partial occlusion. We compare the results with that of standard PCA, employing the Yale face database for the experiments and the results show that ICA is better under certain conditions.

Microcalcification detection in a mammogram is an effective method to find the early stage of breast tumor. Especially, computer aided diagnosis (CAD) improves the working performance of radiologists and doctors as it offers an efficient microcalcification detection. In this paper, we propose a microcalcification detection system which consists of three modules; coarse detection, clustering, and fine detection module. The coarse detection module finds candidate pixels from an entire mammogram which are suspected as a part of a microcalcification. The module not only extracts two median contrast features and two contrast-to-noise ratio features, but also categorizes the candidate pixels with a linear kernel-based SVM classifier. Then, the clustering module forms the candidate pixels into regions of interest (ROI) using a region growing algorithm. The objective of the fine detection module is to decide whether the corresponding region classifies as a microcalcification or not. Eleven features including distribution, variance, gradient, and various edge components are extracted from the clustered ROIs and are fed into a radial basis function-based SVM classifier to determine the microcalcification. In order to verify the effectiveness of the proposed microcalcification detection system, the experiments are performed with full-field digital mammogram (FFDM). We also compare its detection performance with an ANN-based detection system.

A novel edge detector has been developed that utilises statistical masks and neural networks for the optimal detection of edges over a wide range of image types. The failure of many common edge detection techniques has been observed when analysing concealed weapons X-ray images, biomedical images or images with significant levels of noise, clutter or texture. This novel technique is based on a statistical edge detection filter that uses a range of two-sample statistical tests to evaluate any local image texture differences and by applying a pixel region mask (or kernel) to the image analyse the statistical properties of that region. The range and type of tests has been greatly expanded from the previous work of Bowring et al.¹ This process is further enhanced by applying combined multiple scale pixel masks and multiple statistical tests, to Artificial Neural Networks (ANN) trained to classify different edge types. Through the use of Artificial Neural Networks (ANN) we can combine the output results of several statistical mask scales into one detector. Furthermore we can allow the combination of several two sample statistical tests of varying properties (for example; mean based, variance based and distribution based). This combination of both scales and tests allows the optimal response from a variety of statistical masks. From this we can produce the optimum edge detection output for a wide variety of images, and the results of this are presented.