Distinguishing objects on the basis of color is fundamental to humans. In this paper, a clustering approach is used to segment color images. Clustering is usually done using a single point or vector as a cluster prototype. The data can be clustered in the input or feature space where the feature space is some nonlinear transformation of the input space. The idea of kernel principal component analysis (KPCA) was introduced to align data along principal components in the kernel or feature space. KPCA is a nonlinear transformation of the input data that finds the eigenvectors along which this data has maximum information content (or variation). The principal components resulting from KPCA are nonlinear in the input space and represent principal curves. This is a necessary step as colors in RGB are not linearly correlated especially considering illumination effects such as shading or highlights. The performance of the k-means (Euclidean distance-based) and Mixture of Principal Components (vector angle-based) algorithms are analyzed in the context of the input space and the feature space obtained using KPCA. Results are presented on a color image segmentation task. The results are discussed and further extensions are suggested.

Super-resolution based on sequences of low-resolution images has many applications. Among these is improving the image quality of video images, particularly images of historical interest and images from security cameras. Successive frames have slightly different views, or projections, of the object. Not unlike the methods used in computerized tomography, these projections can be combined to produce an image with better resolution than any of the low-resolution views. We observe that in real images even the simplest objects are warped in successive frames. We estimate the warping parameters of each frame and then estimate the object by iterative deconvolution. This forces an appropriate match between a model for the data and the actual data. We show computer simulations of the method and we show some experimental results.

Image segmentation is a process to extract and organize information energy in the image pixel space according to a prescribed feature set. It is often a key preprocess in automatic target recognition (ATR) algorithms. In many cases, the performance of image segmentation algorithms will have significant impact on the performance of ATR algorithms. Due to the variations in feature set definitions and the innovations in the segmentation processes, there is large number of image segmentation algorithms existing in ATR world. Recently, the authors have investigated a number of measures to evaluate the performance of segmentation algorithms, such as Percentage Pixels Same (pps), Partial Directed Hausdorff (pdh) and Complex Inner Product (cip). In the research, we found that the combination of the three measures shows effectiveness in the evaluation of segmentation algorithms against truth data (human master segmentation). However, we still don't know what are the impact of those measures in the performance of ATR algorithms that are commonly measured by Probability of detection (P_Det), Probability of false alarm (P_FA), Probability of identification (P_ID), etc. In all practical situations, ATR boxes are implemented without human observer in the loop. The performance of synthetic aperture radar (SAR) image segmentation should be evaluated in the context of ATR rather than human observers. This research establishes a segmentation algorithm evaluation suite involving segmentation algorithm performance measures as well as the ATR algorithm performance measures. It provides a practical quantitative evaluation method to judge which SAR image segmentation algorithm is the best for a particular ATR application. The results are tabulated based on some baseline ATR algorithms and a typical image segmentation algorithm used in ATR applications.

With recent advances in bandwidth, sensor resolution, and UAV technology, image data is being collected in large quantities. A fast, automated, accurate method to register images is needed because human analysis of this data is time consuming and inaccurate. Once registered, images can be utilized more effectively. Applications where image registration algorithms are used include super-resolution, target recognition, and computer vision. Recent research involved registering images with translation and rotation differences using one iteration of the redundant discrete wavelet transform (rDWT). We extend this work by creating a new multiscale transform to register images with translation or rotation differences. Our two-dimensional multiscale transform uses lowpass filtering and the continuous wavelet transform (CWT) to mimic the two-dimensional rDWT, providing subbands at various scales while maintaining the desirable properties of the rDWT. Our multiscale transform produces data at integer scales, whereas the rDWT produces results only at dyadic scales. We also impose exclusion zones to create spatial separation between significant coefficients. This added flexibility improves registration accuracy without greatly increasing computational complexity and permits accurate registration. Our algorithm's performance is demonstrated by registering test images at various rotations and translations, in the presence of additive white noise.

The quality of satellite images propagating through the atmosphere is affected by phenomena such as scattering and absorption of the light, and turbulence, which degrade the image by blurring it and reducing its contrast. Here, a new approach for digital restoration of Landsat thematic mapper (TM) imagery is presented by implementing several filters as atmospheric filters which correct for turbulence blur, aerosol blur, and path radiance simultaneously. Aerosol modulation transfer function (MTF) is consistent with optical depth. Turbulence MTF is calculated from meteorological data. The product of the two yields atmospheric MTF, which is implemented in the atmospheric filters. Restoration improves both smallness of size of resolvable detail and contrast. Restorations are quite apparent even under clear weather conditions. Different restoration results are obtained by trying to restore the degraded image. Here, restorations results of the Kalman filter and the atmospheric Wiener filter are presented along with restoration results based on wavelets and multifractals. A way to determine which is the best restoration result and how good is the restored image is presented by a visual comparison and by examining several mathematical criteria.

The restoration of vibration-blurred images using measured motion function is considered. Since the blurring filter is of finite impulse response type, the inverse one is of all-pole infinite impulse response type. Direct application of the inverse filter to restore images blurred by vibration is attractive because of minimal computation requirements. However, a pure inverse filter provides excessive noise amplification and is possibly unstable. The proposed technique is to construct a modified inverse filter with preserved all-pole structure and optimized noise and stability properties. An experiment testing the proposed technique was set up and the results presented here.

The Karhunen-Loeve transform based on calculation of the eigenvalues and eigenfunctions of the Karhunen-Loeve integral equation is known to have certain properties which make it optimal for many signal detection and filtering applications. We propose an analytical solution of the equation for a practical case when the covariance function of a stationary process is exponentially oscillating. Computer simulation results using a real aerial image are provided and discussed.

We describe a system for interactive classification and retrieval of microscopic tissue images. Our system models tissues in pixel, region and image levels. Pixel level features are generated using unsupervised clustering of color and texture values. Region level features include shape information and statistics of pixel level feature values. Image level features include statistics and spatial relationships of regions. To reduce the gap between low-level features and high-level expert knowledge, we define the concept of prototype regions. The system learns the prototype regions in an image collection using model-based clustering and density estimation. Different tissue types are modeled using spatial relationships of these regions. Spatial relationships are represented by fuzzy membership functions. The system automatically selects significant relationships from training data and builds models which can also be updated using user relevance feedback. A Bayesian framework is used to classify tissues based on these models. Preliminary experiments show that the spatial relationship models we developed provide a flexible and powerful framework for classification and retrieval of tissue images.

NASA spacecraft have now returned many thousands of images of the surface of Mars. It is no longer practical to analyze such a large dataset by hand, while the development of handwritten feature extraction tools is expensive and laborious. This project investigates the application of machine learning techniques to problems of feature extraction and digital image processing within the Mars dataset. The Los Alamos GENIE machine learning software system uses a genetic algorithm to assemble feature extraction tools from low-level image operators. Each generated tool is evaluated against training data provided by the user. The best tools in each generation are allowed to "reproduce" to produce the next generation, and the population of tools evolves until it converges to a solution or reaches a level of performance specified by the user. Craters are one of the most scientifically interesting and most numerous features on Mars, and present a wide range of shapes at many spatial scales. We now describe results on development of crater finder algorithms using voting sets of simple classifiers evolved by a machine learning/genetic programming system (the Los Alamos GENIE software).

For many image classification applications it is not adequate to take a simple feature extraction approach. A non-parametric approach directly applied to an image can easily result in a problem with a dimension, which can exceed 2^18. Practical application of a neural network classifier requires that some reduction of the image dimension be done prior to classification. This paper compares the performance of several approaches to the problem. Use of wavelets, principal components, and "image zones" are explored in various combinations. The techniques are compared on a specific biometric application.

We propose a new class of generalized moment functions (GMFs) that scan the object with different probing functions. Using the GMF, it is possible to extract a unique geometric point within the object, called the generalized centroid (G-centroid). We can obtain a set of discrete G-centroids from the same object by using different GMFs. The GMFs, which are similar to traditional moment functions, can also be used to describe the global shape of the object, including symmetry and fullness. However, the GMFs, along with the G-centroids, can further serve to construct a feature vector of the object, which is critical to the process of image registration and pattern recognition. Conformal transforms (C-transforms) are another tool used to probe the object by rearranging the latter's mass distribution, without distorting its shape. Using the C-transformed object, it is possible to detect a new mass centroid and G-centroid. More distinguishing feature points can be extracted from the same object by changing the combination of different GMFs and C-transforms. As GMF and centroid detection can be performed by convolution, the centroid and G-centroid can be detected optically in real time. It is also possible to optically implement some of the C-transforms. We present the results of GMF and C-transform applications, including image registration and pattern recognition.

Affected by time difference, sensor and other unsure factors, there is often strip interfering appeared in bi-directional scanning satellite image. The usual method used to remove strip interfering is filtering in frequency field. According to the number of detecting units in detector, using properties of multi-scale and localization, a proper sub-image corresponding to strip noise is got by wavelet decomposition. By processing the sub-image, ideal result is obtained. As far as removing dragging-part and remaining information are concerned, this method is better than the one based on Fourier transform. Experiments with CBERS-1 images have shown that this method is very effective.

Most of the present digital images processing methods are related with objective characterization of external properties as shape, form or colour. This information concerns objective characteristics of different bodies and is applied to extract details to perform several different tasks. But in some occasions, some other type of information is needed. This is the case when the image processing system is going to be applied to some operation related with living bodies. In this case, some other type of object information may be useful. As a matter of fact, it may give additional knowledge about its subjective properties. Some of these properties are object symmetry, parallelism between lines and the feeling of size. These types of properties concerns more to internal sensations of living beings when they are related with their environment than to the objective information obtained by artificial systems. This paper presents an elemental system able to detect some of the above-mentioned parameters. A first mathematical model to analyze these situations is reported. This theoretical model will give the possibility to implement a simple working system. The basis of this system is the use of optical logic cells, previously employed in optical computing.

Diatoms are a large group of microscopic algae that grow as independent organisms or in small colonies and are characterized by a silica cell wall. They are widely used as ecosystem health and evolution tracers. One of the their most important characteristic is the fact that they vary their shape among different families and this makes possible its classification using pattern recognition techniques. In this paper, we describe a new process based on determining the invariant moments of the objects presented in a digital image to classify them according to the different families. We present a technique to segment the diatoms using a new thresholding technique, contour tracing and a recognition process. Finally, the results are presented.

This paper presents a new application for the analysis of histological sections and their 3D visualization. The process is performed in few steps. First, a manual process is necessary to determine the regions of interest, including image digitalization, drawing of borders and alignment between all images. Then, a reconstruction process is made. After sampling the contour, the structure of interest is displayed. The application is experimentally validated and some results on histological sections of a rodent's brain (hamster and rat) are shown.

This paper presents two image techniques for morphometric evaluation. The first one improve the color contrast employing color equalization and borders are identified by using splines. The second one is a semiautomatic method that use fuzzy color thresholding. The second technique will provide the basis of a future automatic method. These techniques are experimentally validated measuring neoformed vessels on histological sections of mice's thigh.

Watermarking techniques have been considerably improved for the last past years, aiming at being always more resistant to attacks. In fact, if the main goal of watermarking at the beginning was to secure digital data (audio, image and video), numerous attacks are still now able to cast doubts on the owner's authenticity ; we can distinguish three different groups of attacks : these one which consist to remove the watermark, these one which aim at impairing the data sufficiently to falsify the detection, and finally these one which try to alter the detection process so that another person becomes the apparent owner of the data. By considering the growing development of always more efficient attacks, this paper firstly presents a recent and exhaustive review of attacks in image and video watermarking. In a second part, the consequences of still image watermarking attacks on video sequences will be outlined and a particular attention will be given to the recently created benchmarks : Stirmark, the benchmark proposed by the University of Geneva Vision Group, this one proposed by the Department of Informatics of the University of Thessaloniki and finally we will speak of the current work of the European Project Certimark ; we will present a comparison of these various benchmarks and show how difficult it is to develop a self-sufficient benchmark, especially because of the complexity of intentional attacks.

This paper deals with some detection issues of watermark signals. We propose an easy way to implement an informed watermarking embedder whatever the detection function. This method shows that a linear detection function is not suitable for side information. This is the reason why we build a family of non-linear functions based on n^th-order statistics. Used with a side-informed embedder, its performance is much better than the classical direct sequence spread spectrum method.

Due to massive development of 3D meshes exchanges on the Internet, transmission and protection of such data has recently received a special focus. In this paper we present a method to compress and watermark the geometry of a 3D triangle mesh in the mesh spectral domain. We first explain how to improve the quality of the compression using overlapping, then we propose a substitutive watermarking scheme in the mesh spectral domain. Finally, we give an overview of the results obtained both for compression or watermarking. The robustness of the watermarking algorithm against the compression method is demonstrated.

A challenging aspect in the development of robust watermarking algorithms is the ability to withstand complex geometric distortion of the media. A few existing techniques are known to deal with such transformations up to a certain level. Traditionally, the measure of the degradation caused by an attack on an image only addressed the pixel value modification. However, a degradation consequent to the geometric distortion of an image cannot be measured with traditional criteria. Therefore the evaluation and comparison of the robustness to desynchronization of different watermarking schemes was not possible. In this paper, we present an innovative method to measure the distortion introduced by complex geometric deformations of an image. The distortion measure is expressed in term of how closely the applied transform can be approximated by a simpler transform model (e.g. RST transform, affine transform). The scheme relies on the local least square estimation of the parameters of the reference transform model. Eventually, we illustrate the proposed measure by presenting some results for different complex image distortions.

This paper proposes a method to embed information into a 3D model represented by a polygonal mesh. The approach used consists in slightly changing the position of the vertices, influencing the length of approximation of the normals to the surface. This technique exhibits relatively low complexity, and offers robustness to simple geometric tranformations. In addition, it does not introduce any visible distortion to the original model.

This paper addresses the problem of controlling the bit rate for image sequences compressed using the Motion JPEG2000 Standard. We propose a computationally efficient iterative technique that is intended for applications where real time (or near real time) encoding is required. Using real world video sequences, we analyze the rate control accuracy and image quality performance of the proposed technique. Although the effectiveness of the technique was demonstrated on high action video sequences, the proposed technique is also applicable to other video sequence encoding applications such as security and surveillance systems or video over the internet.

The performances of discrete-cosine-transform (DCT) JPEG and wavelet-transform (WT) JPEG-2000 for the Karhunen-Loeve-Transform (KLT) based lossy compression of multispectral imagery data are evaluated and compared. The evaluation is based on the measured amount of compression-induced root mean square error in the reconstructed imagery and, more importantly, the impact of compression on the classification of imagery data. We have opted to use classification to assess the impact on compression since it is one of the most widely used forms of machine exploitation procedures. An unsupervised classification via a thematic map is implemented. It is assumed that results for a supervised classification would be similar. The impact of compression is examined at various compression ratios for data obtained from two sensor platforms, LANDSAT TM satellite test imagery with a 30m footprint, and ERIM M7 Sensor aerial test imagery with a 4-6m footprint. Preliminary results, based on the selected test imagery and the selected multispectral bandwidth compression scheme, indicate that the JPEG 2000 generally outperforms the baseline JPEG by a small margin. The results are based on the root-mean-square (RMS) error and the classification accuracy and pertain to imagery with less than 50m footprints. For the 4-6m-footprint ERIM aerial test imagery, JPEG 2000 produces up to four percent higher classification accuracy while incurring up to twelve percent smaller RMS error. However, for the 30m-footprint LANDSAT test imagery, the performance of JPEG and JPEG 2000 are nearly the same. This study does not include imagery with greater than 50m footprint, e.g., NOAA's AVHRR with 1.1 km footprint. For this type of imagery, classification should be performed via a spectral unmixing procedure, instead of a thematic map, since the pixels do not represent pure species.

Error Resilient and Error Concealment 3-D SPIHT (ERC-SPIHT) is a joint source channel coder developed to improve the overall performance against channel bit errors without requiring automatic-repeat-request (ARQ). The objective of this research is to test and validate the properties of two competing video compression algorithms in a wireless environment. The property focused on is error resiliency to the noise inherent in wireless data communication. ERC-SPIHT and MPEG-2 with forward error correction (FEC) are currently undergoing tests over a satellite communication link. The initial test indicates that ERC-SPIHT gives excellent results in noisy channel conditions is shown to have superior performance over MPEG-2 with FEC when communicated over a military satellite channel.

We give a ten-minute tour of scalable video coding technologies, especially those being developed under the auspices of the ISO MPEG. MPEG-4 has already accepted Fine Granularity Scalability (FGS) tools within the Streaming Video Profile. New lines of investigation are underway that include scalable coding based on the emerging Advanced Video Codec (AVC) or H.26L, called AFGS, which significantly outperforms traditional FGS based on MPEG-4, Part 2. A second new line of investigation is in developing fully embedded video coding using motion-compensated 3D wavelet coding, called Interframe Wavelet Coding. In this paper, we present a quick shapshot of these developments, highlighting new results. The encouraging new results indicate that scalable coding may be coming of age.

We give a brief tour of the emerging joint ITU-T/H.264 | ISO/IEC MPEG-4, Part 10 ("AVC") video coding standard. Like its predecessors within ITU-T Recommendations, the draft H.264 standard was originally intended for videoconferencing applications; however, its superior performance over other contenders at a variety of rates and resolutions has led to its adoption by MPEG as the next generation video codec, MPEG-4, Part 10 (dubbed Advanced Video Codec - AVC). This powerful new codec design trades backward compatibility and increased encoder complexity for forward-looking performance and scope of applications. In this paper, we give a status report on this emerging standard, review elements of its current design, and provide some comparisons with the existing standards, MPEG-2, H.263, and MPEG-4 Advanced Simple Profile. Our analysis indicates that the draft H.264 standard offers compelling advantages over all existing video coding standards. It has the potential to redraw the landscape of consumer and enterprise video applications. It reached Final Committee Draft (FCD) status on July 26, 2002.

Novel temporal-spatial rate control solutions for MPEG video transcoding are investigated in this paper. We present two rate control approaches for MPEG video transcoding, one for IPP (including only I and P frames, no B frames) streams and the other for IBP (including I, P and B frames) streams. The proposed rate control schemes adapt the frame rate and the quantization step size simultaneously according to the available channel bandwidth to achieve a good temporal-spatial quality tradeoff. In our proposed solutions, key frames are not allowed for skipping in order to maintain the prediction sequential order, and the MPEG-4 quadratic Rate-Qtz model is adopted to calculate the quantization step size from a given target bit rate. Simulations show that the proposed rate control method for IPP stream transcoding achieves a higher average PSNR value than the MPEG-4 rate control (frame level) with a constant frame rate in CBR channels. Furthermore, the proposed rate control method for IBP stream transcoding can significantly enhance the transcoded video quality in VBR channels, in comparison with straightforward MPEG video transcoding by adjusting the quantization step size at a constant frame rate. The two proposed transcoders are in low computational complexity so that they can be used in real-time transcoding applications.

In media applications there is a high level of available thread level parallelism (TLP). In this paper we study the intra TLP in a video encoder. We show that a well-distributed highly optimized encoder running on a symmetric multiprocessor (SMP) system can run 3.2 faster on a 4-way SMP machine than on a single processor. The multithreaded encoder running on an SMP system is then used to understand the requirements of a chip multiprocessor (CMP) architecture, which is one possible architectural direction to better exploit TLP. In the framework of this study, we use a software approach to evaluate the dataflow between processors for the video encoder running on an SMP system. An estimation of the dataflow is done with L2 cache miss event counters using Intel® VTune^TM performance analyzer. The experimental measurements are compared to theoretical results.

In this paper, a new classification scheme using concentric circles is derived for Vector Quantization (VQ) coder. Two DCT coefficients of each image block are used to compute the radius (r) and angle (θ) in the polar coordinate. A classification scheme consisting of a total of 65 classes is implemented according to the values of r and θ, and is called 65-classes Concentric Circle Classified VQ (65-CCCVQ). Simulation results show that using 65-CCCVQ, the compression speed is much faster than the original VQ algorithm while maintaining similar image quality.

The Consultative Committee for Space Data Systems (CCSDS) has been engaging in recommending data compression standards for space applications. The first effort focused on a lossless scheme that was adopted in 1997. Since then, space missions benefiting from this recommendation range from deep space probes to near Earth observatories. The cost savings result not only from reduced onboard storage and reduced bandwidth, but also in ground archive of mission data. In many instances, this recommendation also enables more science data to be collected for added scientific value. Since 1998, the compression sub-panel of CCSDS has been investigating lossy image compression schemes and is currently working towards a common solution for a single recommendation. The recommendation will fulfill the requirements for remote sensing conducted on space platforms.

Eastman Kodak Company has continued the development of a rate-controlled, adaptive Differential Pulse Code Modulation (DPCM) image compression algorithm for commercial remote sensing applications. A version of this algorithm is currently being used in a space-qualified ASIC onboard the Space Imaging IKONOS satellite and Digital Globe QuickBird satellite. This ASIC compresses the raw imagery data (before calibration) at a speed just under 4 megapixels per second. Kodak has redesigned this ASIC to increase the functionality and throughput, while maintaining the power and reducing the size. This new ASIC is intended to meet the future commercial remote sensing requirements for increased resolution and greater area coverage.

The present contribution reports the development of an application specific integrated circuit (ASIC) for application in space-borne electronic data handling systems. The function of the ASIC is to perform a two-dimensional, multi-level Discrete Wavelet Transform (DWT) at high data rate. The implemented DWT is detailed mathematically, and the functionality and the architecture of the ASIC are described. The DWT ASIC will perform pixel data decorrelation in the CWIC on-board image compressor.

Thanks to its intrinsic scalability features, the wavelet transform has become increasingly popular as decorrelator in image compression applications. Throuhgput, memory requirements and complexity are important parameters when developing hardware image compression modules. An implementation of the classical, global wavelet transform requires large memory sizes and implies a large latency between the availability of the input image and the production of minimal data entities for entropy coding. Image tiling methods, as proposed by JPEG2000, reduce the memory sizes and the latency, but inevitably introduce image artefacts. The Local Wavelet Transform (LWT), presented in this paper, is a low-complexity wavelet transform architecture using a block-based processing that results in the same transformed images as those obtained by the global wavelet transform. The architecture minimizes the processing latency with a limited amount of memory. Moreover, as the LWT is an instruction-based custom processor, it can be programmed for specific tasks, such as push-broom processing of infinite-length satelite images. The features of the LWT makes it appropriate for use in space image compression, where high throughput, low memory sizes, low complexity, low power and push-broom processing are important requirements.

Future high resolution instruments planned by CNES for space remote sensing missions will lead to higher bit rates because of the increase in both resolution and number of bits per pixel, not compensated by the reduced swath. Data compression is then needed, with compression ratio goals always higher and with artifacts remaining unnoticeable. Up to now studied algorithms are based on intra-band coding and utilize the intra-image or spatial correlation. The spaceborne earth observation instruments have however several spectral channels (one panchromatic band and at least 3 spectral bands) and since such algorithms process independently each channel, the inter-image or spectral correlation is ignored. For optimum compression performance, multispectral algorithms have to be studied in order to exploit both spectral and spatial correlation. This paper proposes a low complexity and flexible fixed data rate compression algorithm for multispectral imagery.

Analysis and understanding of medical images has important clinical values for patient diagnosis and treatment, as well as technical implications for computer vision and pattern recognition. One of the most fundamental issues is the detection of object boundaries or singularities, which is often the basis for further processes such as organ/tissue recognition, image registration, motion analysis, measurement of anatomical and physiological parameters, etc. The focus of this work involved taking a correlation based approach toward edge detection, by exploiting some of desirable properties of wavelet analysis. This leads to the possibility of constructing a bank of detectors, consisting of multiple wavelet basis functions of different scales which are optimal for specific types of edges, in order to optimally detect all the edges in an image. Our work involved developing a set of wavelet functions which matches the shape of the ramp and pulse edges. The matching algorithm used focuses on matching the edges in the frequency domain. It was proven that this technique could create matching wavelets applicable at all scales. Results have shown that matching wavelets can be obtained for the pulse edge while the ramp edge requires another matching algorithm.

JPEG 2000 is the emerging image compression standard. Part 1 of the standard was approved in January 2001, and software implementations of JPEG 2000 are starting to appear in the industry. JPEG 2000 might rapidly become one of the more commonly used compressed image file formats. What then would be an ideal way to showcase the rich feature set that JPEG 2000 offers? In this paper we describe how JPEG 2000 technology was enabled on a handheld device, including the manner by which it was designed and implemented, the software and hardware tools that were used for the development, the target PDA (Personal Digital Assistant) and operating system, the roadblocks that were encountered along the way and subsequently overcome, and the results and feedback for the completed application, which was demonstrated on the floor at Comdex 2001.

The Parallel Computational Environment for Imaging Science, PiCEIS is an image processing code designed for efficient execution on massively parallel computers. Through effective use of the aggregate resources of such computers PiCEIS enables much larger and more accurate production processing using existing off the shelf hardware. A goal of PiCEIS is to decrease the difficulty of writing scalable parallel programs and reduce the time to add new functionalities. In part this is accomplished by the PiCEIS architecture, its ability to easily add additional modules, and also through the use of a non-uniform memory access (NUMA) programming model based upon one-sided access to distributed shared memory. In this paper we briefly describe the PiCEIS architecture, our NUMA programming tools, and examine some typical techniques and algorithms.

We have developed high density image processing techniques for finding the surface strain of an untreated sample of material from a sequence of images taken during the application of force from a test rig. Not all motion detection algorithms have suitable functional characteristics for this task, as image sequences are characterised by both short- and long-range displacements, non-rigid deformations, as well as a low signal-to-noise ratio and methodological artifacts. We show how a probability-based motion detection algorithm can be used as a high confidence estimator of the strain tensor characterising the deformation of the material. An important issue discussed is how to minimise the number of image brightness differences that need to be calculated. We give results from two studies of materials under axial tension: a sample of aluminium alloy exhibiting a propagating plastic deformation, and a preparation of deer antler bone, a natural composite material.

This paper presents a Landweber-type iterative demosaicking method for a single solid state color image sensor using three primary color filters of the Bayeris pattern. The iterative demosaicking method can restore a high-resolution color image much better than the existing non-iterative demosaicking methods using switchover-type interpolation filters. This paper, extending the idea of our previously presented image acquisition scheme using an imperfect image sensor with defect pixels, forms the iterative demosaicking method. Our previous scheme prepares a defect-pixel map in advance, and then takes a defocused image with an imperfect image sensor. Utilizing the defect-pixel map and the information that each pixel shares with its adjacent pixels, it recovers defect pixels and simultaneously sharpens the blurred image. As the recovery technique, we have formed the Landweber-type iterative algorithm. Taking it into account that decimated color pixels caused by the color filters correspond to defect pixels in the above recovery problem, we adapt our previous iterative recovery method to the demosaicking problem. Furthermore, to restrain the occurrence of false color caused by the demosaicking and simultaneously to preserve original hue variations as thoroughly as possible, this paper forms a hybrid demosaicking method that first applies the Landweber-type iterative algorithm to the interpolation of the green color component and then performs the interpolation of the red and the blue components with some existing hue- or chrominance-preserving-type method. Experiments using test images and color images really taken with a high-resolution digital color camera demonstrate that our hybrid demosaicking method reproduces a sharpened high-resolution color image without producing noticeable artifacts of false color.

In this paper a new method for a fast out-of-focus blur estimation and restoration is proposed. It is suitable for CFA (Color Filter Array) images acquired by typical CCD/CMOS sensor. The method is based on the analysis of a single image and consists of two steps: 1) out-of-focus blur estimation via Bayer pattern analysis; 2) image restoration. Blur estimation is based on a block-wise edge detection technique. This edge detection is carried out on the green pixels of the CFA sensor image also called Bayer pattern. Once the blur level has been estimated the image is restored through the application of a new inverse filtering technique. This algorithm gives sharp images reducing ringing and crisping artifact, involving wider region of frequency. Experimental results show the effectiveness of the method, both in subjective and numerical way, by comparison with other techniques found in literature.

Numerous car accidents occur on urban road. However, researches done so far on driving assistance are subjecting highways whose environment is relatively simple and easy to handle, and new approach for urban settings is required. Our purpose is to extend its support to the following conditions in city traffic: the presence of obstacles such as pedestrians and telephone poles; the lane mark is not always drawn on a road; drivers may lack the sense of awareness of the lane mark. We propose a collision avoidance system, which can be applied to both highways and urban traffic environment. In our system, stereo cameras are set in front of a vehicle and the captured images are processed through a computer. We create a Projected Disparity Map (PDM) from stereo image pair, which is a disparity histogram taken along ordinate direction of obtained disparity image. When there is an obstacle in front, we can detect it by finding a peak appeared in the PDM. With a speed meter and a steering sensor, the stop distance and the radius of curvature of the self-vehicle are calculated, in order to set the observation-required area, which does not depend on lane marks, within a PDM. A danger level will be computed from the distance and the relative speed to the closest approaching object detected within the observation-required area. The method has been tested in urban traffic scenes and has shown to be effective for judging dangerous situation, and gives proper alarm to a driver.

Our aim is to prevent rear-end collision of the vehicle, and we propose a new method for a front-vehicle observation system using sequence of road images. In this research, we have developed an in-vehicle system that can be applied to a software-based rear-end collision with low computing power. First, we calculate projective transformation (a vehicle’s front image is changed into the images seen from right above) of the vehicle’s front image acquired by a single camera placed on the dashboard of our vehicle. Next, the difference of two projective transformation images is calculated. Finally, we calculate the degree of crashing risks depending on the distance to the front vehicle and the speed of our vehicle, and a warning is given to the driver. This system mounted on a vehicle has been tested on expressways. As a result, the effectiveness of our method has been verified.

JPEG 2000 is the emerging image compression standard. The base standard (Part 1) was approved in January 2001. Several corporations, research organizations, and individual parties are already offering software implementations of, at minimum, Part 1 of the specification. This paper describes the test metrics and results for some of the currently available JPEG 2000 software implementations. The software implementations are also compared against the JPEG 2000 Verification Model (VM) and the Independent JPEG Group's JPEG implementation (IJG). The evaluation testing is performed under both the Unix (Solaris) and Windows operating systems. Results will be presented from the metrics categories mentioned above.

In order to perform dosage analysis of prostate seed implants, 3D seed coordinates are ascertained for seed positions from conical x-ray projections. A conventional reconstruction approach uses back projection of x-ray data obtained at two or three gantry angles. The method, however, does not perform well when the seeds are obscured by other seeds in a projection. Additional x-ray projection data taken at different angles can resolve the overlapping seeds but the computational overhead increases dramatically. In this paper, we propose an alternate approach for 3D seed localization using Hough transformation. For each seed's coordinate in three dimensions, there exists a corresponding projection location in each of the views. Integrating each of these seed locations and placing the result in the 3D seed coordinate results in a high confidence score if there is a seed at the coordinate. This technique does not suffer from the problem of inability to reconstruct the overlapping seeds since the seed's path is unique for a particular seed. When seeds are overlapped, the paths intersect each other. From the seed's path, the overlapping seeds' coordinates can be determined. Our proposed method also has the ability to incorporate more views without incurring major computing cost. Using Hough transform parametric equations to describe the path of the seeds from one view to the next, the Hough transform weight of the 3D seed coordinates among the views can be calculated. Results from computer simulation and a physical phantom study are presented to illustrate the proposed approach. The results indicate that the Hough transform method can determine the 3D position of a brachytherapy seed even when the seed is undetected in some of the x-ray projections due to the seed overlap.

We present a vector quantization scheme acting on brightness fields based on distance/distortion criteria correspondent with psycho-visual aspects. These criteria quantify sensorial distortion between vectors that represent either portions of a digital image or alternatively, coefficients of a transform-based coding system. In the latter case, we use an image representation model, namely the Hermite transform, that is based on some of the main perceptual characteristics of the human vision system (HVS) and in their response to light stimulus. Energy coding in the brightness domain, determination of local structure, code-book training and local orientation analysis are all obtained by means of the Hermite transform. This paper, for thematic reasons, is divided in four sections. The first one will shortly highlight the importance of having newer and better compression algorithms. This section will also serve to explain briefly the most relevant characteristics of the HVS, advantages and disadvantages related with the behavior of our vision in front of ocular stimulus. The second section shall go through a quick review of vector quantization techniques, focusing their performance on image treatment, as a preview for the image vector quantizer compressor actually constructed in section 5. Third chapter was chosen to concentrate the most important data gathered on brightness models. The building of this so-called brightness maps (quantification of the human perception on the visible objects reflectance), in a bi-dimensional model, will be addressed here. The Hermite transform, a special case of polynomial transforms, and its usefulness, will be treated, in an applicable discrete form, in the fourth chapter. As we have learned from previous works ¹, Hermite transform has showed to be a useful and practical solution to efficiently code the energy within an image block, deciding which kind of quantization is to be used upon them (whether scalar or vector). It will also be a unique tool to structurally classify the image block within a given lattice. This particular operation intends to be one of the main contributions of this work. The fifth section will fuse the proposals derived from the study of the three main topics- addressed in the last sections- in order to propose an image compression model that takes advantage of vector quantizers inside the brightness transformed domain to determine the most important structures, finding the energy distribution inside the Hermite domain. Sixth and last section will show some results obtained while testing the coding-decoding model. The guidelines to evaluate the image compressing performance were the compression ratio, SNR and psycho-visual quality. Some conclusions derived from the research and possible unexplored paths will be shown on this section as well.

Authors have worked out a nonstationary signal analysis method on an example of research laser lines. This method disclosed relationship between signal approximation coefficients and geometry signal characterizations (for instance, energy center, moment of inertia). An example, which is demonstrating an application of this method for exact coordinate determination problem in laser line at displacement compensation in laser imaging are present.

The William Blake Archive is part of an emerging class of electronic projects in the humanities that may be described as hypermedia archives. It provides structured access to high-quality electronic reproductions of rare and often unique primary source materials, in this case the work of poet and painter William Blake. Due to the extensive high frequency content of Blake's paintings (namely, colored engravings), they are not suitable for very efficient compression that meets both rate and distortion criteria at the same time. Resolving that problem, the authors utilized modified Mixed Raster Content (MRC) compression scheme -- originally developed for compression of compound documents -- for the compression of colored engravings. In this paper, for the first time, we have been able to demonstrate the successful use of the MRC compression approach for the compression of colored, engraved images. Additional, but not less important benefits of the MRC image representation for Blake scholars are presented: because the applied segmentation method can essentially lift the color overlay of an impression, it provides the student of Blake the unique opportunity to recreate the underlying copperplate image, model the artist's coloring process, and study them separately.

Wavelet based image compression has been a focus of research in recent days. In this paper, we propose a compression technique based on modification of original EZW coding. In this lossy technique, we try to discard less significant information in the image data in order to achieve further compression with minimal effect on output image quality. The algorithm calculates weight of each subband and finds the subband with minimum weight in every level. This minimum weight subband in each level, that contributes least effect during image reconstruction, undergoes a threshold process to eliminate low-valued data in it. Zerotree coding is done next on the resultant output for compression. Different values of threshold were applied during experiment to see the effect on compression ratio and reconstructed image quality. The proposed method results in further increase in compression ratio with negligible loss in image quality.

There are two kinds of image in the field of imaging for the underwater target due to using two types of light source, respectively, continue green-blue laser and pulse green-blue laser. When pulse laser is used, the obtained image has the characteristic of low contrast, small dimension of target and complex background. The method of multiple correlation peak-value detection, which is based on the correlativity of pixels in the region of target, can be directly used to implement target detection for the underwater laser image. When continue laser is used in the process of imaging for the underwater target, there is a beam of strong light in the image, which makes it difficult to implement target detection. Therefore, we use the multi-scale and multi-channel Gabor filter group with the character of direction selection to depress the disturbance of the strong beam firstly, and then, use the multiple correlation analysis method to detect the weak target in the image. Experimental results show that these methods are practical and effective.

This article is a continuation of our work on signal detection in image sequences ^{1, 2}. An algorithm has been developed for detection, localization and tracking of signals that appear in a small number of pixels of an image sequence and are periodic in the temporal domain. For such a signal, uniformly sampled, a model dependent on few parameters is derived and used for fitting a power spectrum and the signal itself. To enhance such a signal, a generalization of the Principal Component Method is proposed. A simulation based on the model is described. The pre-processing procedure is basically the same as before ^{1, 2}. The algorithm was tested by processing simulated image sequences, as well as some real ones.

The secure identification of parasites can be problematic and yet it is of prime importance. However, the specific identification of the parasites with traditional techniques can be slow and time consuming, requiring quality preparations where each taxonomically important structure can be clearly observed. Color information becomes an important discriminant feature, which we need to include in the whole identification process. Digital images of the monogeneans: Heterobothrium ecuadori and Neobenedenia melleni and, the digeneans Lintonium vibex, Homalometron longisinosum, Bianium plicitum and Phyllodistomum mirandai were processed to obtain their diffraction patterns. A numerical simulation was performed in order to correlate diffraction patterns of parasites species with phase only filters. The position, scale and rotation invariant image recognition was made through the scale transform.

Fine resolution synthetic aperture radar (SAR) and interferometric synthetic aperture radar (IFSAR) have been widely used for the purpose of creating viable terrain maps. A map is only as good as the information it contains. Therefore, it is a major priority of the mapmakers that the data that goes into the process be as complete and accurate as possible. In this paper, we analyze IFSAR correlation/de-correlation data to help in terrain feature information. The correlation data contains the correlation coefficient between the bottom and top IFSAR radar channels. It is a 32-bit floating-point number. This number is a measure of the absolute complex correlation coefficient between the signals that are received in each channel. The range of these numbers in between zero and unity. Unity indicates 100% correlation and zero indicates no correlation. The correlation is a function of several system parameters including signal-to-noise ratio (SNR), local geometry, and scattering mechanism. These two radar channels are physically close together and signals are inherently highly correlated. Significant difference is found beyond the fourth decimal place. We have concentrated our analysis on small features that are easily detectable in the correlation/de-correlation data and not so easily detectable in the elevation or magnitude data.

The taxonomic identification of diatom species that constituted phytoplankton communities in remote times is determined in several research fields like ecology, evolution, paleocology and biostratigraphy. In the last 30 years the use of fossil diatoms like environmental indicators has become of prime importance. However the use of these organisms is limited since they are found in sediment samples mostly fragmented or pulverized. This may lead to confusion and loss of information. In this work we used invariant correlation to identify 21 species of fossil diatoms. This correlation method is invariant to position, scale and rotation of the image. With this method we were able to identify the diatom species from only a small fragment of the organisms. Results showed that it is possible to identify some species having a range since 2.12% of information of the image. For example the minimum percentage was for Azpeitia nodulifer var A. This methodology can be used for the development of an automated system of plankton identification. An automatized identification of diatoms would be able to guarantee a faster identification and also would reduce the time necessary for accomplishing analysis of samples highly fragmengted.

In this paper, the finite-state vector quantizers (FSVQs) with an extended super codebook obtained by applying the affine transformation to the codewords are proposed for the image coding framework. In designing the state codebook, each codeword in the super codebook in conventional FSVQs is affine-transformed (luminance shift, contrast scaling, and isometry operations) and thus a much larger virtual codebook is obtained. By using the matching criteria, such as the gradient match and the side match criteria in existing FSVQs, for the neighboring pixels in the block boundaries the much smaller state codebooks with different sizes can be constructed. Note that the rest parts of the proposed scheme are similar to those in the conventional FSVQs. On the other hand, much higher image quality can be obtained by using the extended virtual codebook. According to our simulation results, the peak-signal-to-noise ratio (PSNR) of the decoded images is significantly improved by 1--2 dB for the extended virtual codebook. However, the PSNRs are significantly reduced once the FSVQ with the gradient match and side match criteria are employed. Therefore we proposed another scheme that can improve the PSNR under the same bit rate. The simulation results show that the PSNR can be slightly increased. According to this scheme, we will propose efficient scheme to significantly increase the PSNR in our future work.

The increase in traffic accidents is becoming a serious social problem with the recent rapid traffic increase. In many cases, the driver’s carelessness is the primary factor of traffic accidents, and the driver assistance system is demanded for supporting driver’s safety. In this research, we propose the new method of automatic detection and recognition of road signs by image processing. The purpose of this research is to prevent accidents caused by driver’s carelessness, and call attention to a driver when the driver violates traffic a regulation. In this research, high accuracy and the efficient sign detecting method are realized by removing unnecessary information except for a road sign from an image, and detect a road sign using shape features. At first, the color information that is not used in road signs is removed from an image. Next, edges except for circular and triangle ones are removed to choose sign shape. In the recognition process, normalized cross correlation operation is carried out to the two-dimensional differentiation pattern of a sign, and the accurate and efficient method for detecting the road sign is realized. Moreover, the real-time operation in a software base was realized by holding down calculation cost, maintaining highly precise sign detection and recognition. Specifically, it becomes specifically possible to process by 0.1 sec(s)/frame using a general-purpose PC (CPU: Pentium4 1.7GHz). As a result of in-vehicle experimentation, our system could process on real time and has confirmed that detection and recognition of a sign could be performed correctly.

In this research, we propose the concept of the “QOD (The Quality Of Driving)”. We have already established the method for quantifying the state of driving, and then estimate the “QOD”. To evaluate the “QOD” of the driving state in actual urban road, we adopted three parameters: the degree of the meandering, the stability of the acceleration, and the sufficiency of the distance between own vehicle and the vehicle ahead. In our system, we measure these parameters from the sequence of images taken by the camera placed on the dashboard of our vehicle, and evaluate the “QOD”. Specifically, we acquire the FOE (Focus Of Expansion) by using the optical flow vectors obtained from the sequence of images, and calculate the yaw rate of our vehicle, to get the amount of meandering. Next, velocity of the car is obtained from a velocity sensor and the stability of the acceleration is computed. Finally, we extract the domain of the self-lane in road images through the Hough Transform and then detect the horizontal edges within the domain. The front vehicle can be detected from the optical flow vectors emerging out from these edges. Then we calculate the distance between own vehicle and the vehicle ahead by using a camera parameter. The distance between two vehicles is computed by considering the relationship between the distance and the velocity.

Fingerprint recognition involves in two main approaches: minutiae-based algorithm, the most popular and traditional, has several drawbacks and not suitable for the applications that using solid-state sensor; while structure-based algorithm, using the Gabor filters, captures rich discriminatory information contained in the gray level fingerprint image and generates a feature code with the same length, which is benefit for quickly matching because of the bit wise comparison. This paper probes into the optimal design of Gabor filters theoretically. The inference of optimal parameters of Gabor filters and the whole identification procedure is described in detail. Many comparison experiments are also considered carefully, such as the size of tessellation cells, etc. The experiment result shows the efficiency of the Gabor filters. And to reduce the identification time, we remain the main filters while maintaining same recognition accuracy.

There is a need for autonomous submarines that can quickly and safely complete jobs, such as the recovery of a downed aircraft's black box recorder. In order to complete this feat, it is necessary to use an optical processing algorithm that distinguishes a desired target and uses the feedback from the algorithm to retrieve the target. The algorithm itself uses many bit mask filters for particle information, and then uses a unique rectation method in order to resolve complete objects. The algorithm has been extensively tested on an AUV platform, and proven to succeed repeatedly in approximately five or more feet of water clarity.

JPEG 2000 is a new image compression standard from ISO/IEC JTC1 SC29 WG1, the Joint Photographic Experts Group (JPEG) committee. Better thought of as a sibling to JPEG rather than descendant, the JPEG 2000 standard offers wavelet based compression as well as companion file formats and related standardized technology. This paper examines the JPEG 2000 standard for features in four specific areas-compression, file formats, client-server, and conformance/compliance that enable image workflows.