An increasing number of applications, such as audio and video conferencing, are using the internet as a communication backbone. In such applications it is important to minimize the latency of communication while extracting sufficient throughput. Existing applications tend to fill the internet with large amounts of data with the result that maximum throughput is achieved only at the cost of very high latency. In the case of packet switched networks there are large delays due to the presence of the routers in the path and due to the networks being shared. We have developed a protocol to control the latency for communications over the internet while holding throughput to an acceptable level and reducing jitter. The scheme is general and adaptive, so it can work with a variety of applications and protocols. An important aspect of our work is that it provides a possible solution for reducing latency in communications over the existing internet without using any additional hardware.

We are developing an intelligent robot camera that is capable of automatically shooting images like those taken by broadcasting cameramen. In order to elucidate the zooming techniques of broadcasting cameramen, we developed the new measurement tool of zooming works which is able to be attached to or detached from broadcasting cameras easily. Then we measured the zooming works of broadcasting cameramen in shooting the images of still subject with this tool, and analyzed them. We confirmed that it is possible to reproduce faithfully the subtle zooming techniques of broadcasting cameramen. It was also established that when a cameraman shoots the same scene repeatedly in an improvised manner, the velocity of the operation rod follows almost the same velocity curve each time. This has led us to believe that the cameramen conduct zooming on the basis of a method acquired through individual experience. The frequency components of the operation rod's subtle velocity changes are from 5 to 10 Hz and the frequency levels of less than 5 Hz are especially large. The peak of variation of the subject's size is observed at the timing of 0.53 to 0.66 (the second half) during zoom-in operation and at the timing of 0.38 to 0.47 (the first half) during zoom-out operation, with both measured in terms of normalized zooming time.

In the present work we develop a biologically inspired image coding system. The distribution of photoreceptors across the human retina is not uniform which provides a space-variant acuity since the sampling rate varies with the visual angle. Such an organization allows to selectively reduce the amount of visual information to process through attention mechanisms. Then, while most current computer vision models work with a constant spatial resolution, we study the use of a space- variant architecture in the image processing context. The method we present here is an illustration of such an approach.

The conventional approach of using DCT (discrete cosine transform) which is Fourier based has been most popular in industry. The advantages and potentials of image compression using wavelet analysis are now being explored. In this paper a new image compression algorithm based on wavelet analysis is presented that makes use of a lifting scheme and a modification of the 3-D subband coding. The exploitation of contexts in both space-time and space-space brings out an improved algorithm that is both performance effective and computationally efficient. The software system using this algorithm has the advantage that it is much more flexible and less costly than hardware systems. The new algorithm has been tested with the compression of traffic scene video with satisfactory results.

This paper treats the compression of Synthetic Aperture Radar (SAR) imagery. SAR images are difficult to compress, relative to natural images, because SAR contains an inherent high frequency speckle. Today's state-of-the-art coders are designed to work with natural images, which have a lower frequency content. Thus, their performance on SAR is under par. In this paper we given an overview performance report on the popular compressions techniques, and investigate three approaches to improve the quality of SAR compression at low- bit rates. First, we look at the design of optimal quantizers which we obtain by training on SAR data. Second, we explore the use of perceptual properties of the human visual system to improve subjective coding quality. Third, we consider the use of a model that separates the SAR image into structural and textural components. The paper concludes with a subjective evaluation of the algorithms based on the CCIR recommendation for the assessment of picture quality.

We propose two methods to provide optimal image quality at a fixed image delivery rate for any given transmission channel condition. The first method, channel-controlled variable-rate (CCVR) image coding, employs adaptive-rate source coding and channel coding, while operating with a fixed modulation symbol rate. The second method, adaptive-rate coding-modulation (ARCM), extends the CCVR system by utilizing adaptive modulation. Both methods use a variable-compression-ratio image coder and variable-rate channel coding. The objective is to maximize the quality of the reconstructed image at the receiver when transmitted through Rayleigh fading and additive white Gaussian noise (AWGN). The CCVR system maximizes the reconstructed image quality through a bit-rate trade-off between the source and channel coders. The ARCM method provides a trade-off between the rates of source and channel coding, and the modulation rate. Both methods require knowledge of the channel state which is used by the receiver to inform the transmitter, via a feedback channel, of the optimal strategy for image compression, channel coding, and modulation format. The resulting system using ARCM achieves up to a 17 dB improvement over the peak signal-to-noise ratio (PSNR) performance of a system using a fixed-compression-ratio image coder and fixed-rate channel coding. We extend the adaptive image transmission system to video, and present a candidate system for wireless mobile video applications using orthogonal frequency division multiplexing (OFDM).

This paper introduces a new scheme for still image compression based on Vector Quantization (VQ). The new scheme first vector quantized the image, then the indices obtained from quantization are compressed and transmitted. The indices are used as a classifier to identify the active areas of the image. The residual of active areas are vector quantized in the second step and the indices generated are transmitted. The advantage of new scheme is to present the active areas of the coded image accurately without overhead requirement. This scheme shows better subjective and objective in comparison with similar VQ schemes.

This paper describes a wavelet domain restoration algorithm, based on a so-called C/D/C system model that accounts for system blur at the image acquisition level, for the potentially important effects of aliasing, and for additive noise. Wavelet domain modeling discretizes the image acquisition kernel, and in this way the image restoration problem is formulated as a discrete least squares problem. The treatment of noise is related to the singular values of the image acquisition kernel. The performance of our wavelet domain restoration algorithm is assessed. We show that pixel- scale features can be restored exactly in the absence of noise. In the presence of noise results are similar, except for some noise-amplification and ringing artifacts that we control with the proper choice of an algorithm parameter. This paper extends work in wavelet domain C/D/C model-based restoration, and builds on research using the C/D/C model in digital image restoration.

We advance active computer vision algorithms for flexible manufacturing systems that classify objects and estimate their pose from intensity images. Our algorithms automatically reposition the sensor if the class or pose of an object is ambiguous in a given image and incorporate data from multiple object views in determining the final object classification. A feature space trajectory (FST) in a global eigenfeature space is used to represent 3-D distorted views of an object. Bayesian methods are used to derive the class hypothesis, pose estimate, confidence measures, and the sensor position that best resolves ambiguity. FSTs constructed using images rendered from solid models of objects are used to process real image data.

A unified approach to image focus and defocus analysis (UFDA) was proposed recently for three-dimensional shape and focused image recovery of objects. One version of this approach which yields very accurate results is highly computationally intensive. In this paper we present a parallel implementation of this version of UFDA on the Parallel Virtual Machine (PVM). One of the most computationally intensive parts of the UFDA approach is the estimation of image data that would be recorded by a camera for a given solution for 3D shape and focused image. This computational step has to be repeated once during each iteration of the optimization algorithm. Therefore this step has been sped up by using the Parallel Virtual Machine (PVM). PVM is a software package that allows a heterogeneous network of parallel and serial computers to appear as a single concurrent computational resource. In our experimental environment PVM is installed on four UNIX workstations communicating over Ethernet to exploit parallel processing capability. Experimental results show that the communication over-head in this case is relatively low. An average of 1.92 speedup is attained by the parallel UFDA algorithm running on 2 PVM connected computers compared to the execution time of sequential processing. By applying the UFDA algorithm on 4 PVM connected machines an average of 3.44 speedup is reached. This demonstrates a practical application of PVM to 3D machine vision.

This paper introduces a signal processing strategy for depth segmentation and scene reconstruction that incorporates occlusion as a natural component. The work aims to maximize the use of connectivity in the temporal domain as much as possible under the condition that the scene is static and that the camera motion is known. An object behind the foreground is reconstructed using the fact that different parts of the object have been seen in different images in the sequence. One of the main ideas in this paper is the use of a spatio- temporal certainty volume c(x) with the same dimension as the input spatio-temporal volume s(x), and then use c(x) as a 'blackboard' for rejecting already segmented image structures. The segmentation starts with searching for image structures in the foreground, eliminate their occluding influence, and then proceed. Normalized convolution, which is a Weighted Least Mean Square technique for filtering data with varying spatial reliability, is used for all filtering. High spatial resolution near object borders is achieved and only neighboring structures with similar depth supports each other.

In this paper we present three image processings intended to limit eye strain when using stereoscopic displays. These processings are based on a compatibility of the disparity in the image and of the local spatial frequency: to measure this compatibility, we introduced a stereoscopic comfort function. According to the value of the function, one point in the image is processed or not. We suggest three filterings (a 'Virtual Curtain,' a 'Virtual Pane' and an 'Adaptive Haze') that hide the points behind a gray curtain, a frosted pane or a kind of aerial perspective haze. These three processings were applied to synthetic images and presented to subjects in a human factors experiment. Whereas the less flexible processings ('Virtual Curtain' and 'Virtual Pane') were not very convincing, the 'Adaptive Haze' proved to be more comfortable than the two others, and even more than the original images to a large extent.

Inspired by models of natural vision processing, intensity and color information are encoded and successfully decoded using coarse coding. Primate photoreceptors are known to include one rod and three cones, each with a unique spectral absorption curve. Although the curves overlap significantly, vision systems are capable of incredible chromatic acuity and spatial luminous acuity. A proof of concept is demonstrated here using simulated absorption curves and an algorithm representing a cursory model of vision processing. The color and intensity of objects within test images are successfully retrieved after passing through only two coarse filters arranged in a checkerboard fashion. A consequence of such a filter is a natural edge enhancement of the objects within the image.

In this paper, we present a supervised, nonparametric Bayesian approach to object detection problem in color images. The proposed method does not assume any a priori knowledge about the object color distributions and their forms. It estimates the underlying color densities directly from the given image data. In this respect, the algorithm is not affected by substantial variations in the input scenes. For density estimation, the object color distributions are expanded in a series of orthogonal basis functions in the (R,G,B)-color space. The basis functions selected are the Hermite polynomials since their region of orthogonality spans the entire red, green, and blue coordinate axes. Several color images of complex shaped objects are processed by the detection method in the computer simulation. The input images are varied in size to determine the computational cost of the algorithm. During the training phase, the object and background color densities of the selected images are estimated using the first seventeen Hermite polynomials. In the classification phase, the nonparametric Bayesian classifier is applied to the input images to detect the objects of interest. Even for noisy images of textured scenes, the algorithm has detected the regions of interest with high accuracy. The experimental results indicate that the seventeen lowest-order Hermite polynomials are sufficient for accurate color density estimation. The computational cost of the method is also reasonably low for this type of iterative density estimation process.

Image enhancement applications are highly dependent on the efficiency of edge detection techniques. Most of these techniques have a time complexity of O(n²) where the picture has size n X n. The use of more advanced algorithms can substantially reduce this requirement, improving the computational performance of the application. This paper presents a new method, named GALE, which combines the random search mechanisms of Genetic Algorithms with linear time methods. The resulting edge detection process approaches linear time complexity as demonstrated in the experiments also reported here. The Genetic Algorithm is constructed by utilizing a fitness measurement which is proportional to a directional gradient to select picture windows and establishes candidate pairs of points which bracket an edge. Such areas are then investigated by using near-neighbor linear techniques and the Sobel number for edge identification and detection. The linear technique procedures are built in such a way that the use of other fitness functions, such as the Sombrero operator, instead of the Sobel number are easily implemented and activated. The paper begins by discussing related work in this area, following by the description of the basic concepts of Genetic Algorithms required for this solution. A detailed view of the linear search algorithm is then presented, followed by a report on some experiments conducted in a controlled environment. Theoretical results are used to support the evidence of the time complexity and correctness of this new method. In addition, the experimental results show the improved performance of this method.

Edges or perceptible intensity transitions in digital imagery are identified from the zero-crossings of Laplacian of Gaussian (LOG) filtered images. Time or frequency-sampled LOG filters have been developed for the detection of edges in digital image data. The image is decomposed into overlapping subblocks and processed in the transform domain. In order to achieve accurate and efficient implementations, the discrete symmetric cosine transform (DSCT) of the input data is employed in conjunction with adaptive filters. The adaptive selection of the filter coefficients is based on the gradient criterion. For instance, in the case of the frequency-sampled LOG filter, the filter parameter is systematically varied to force the rejection of spurious edge classifications. In addition, the proposed algorithms easily extend to higher dimensions. This is useful where 3-D medical image data containing edge information has been corrupted by noise. This paper employs isotropic and non-isotropic filters to track edges in such images. The algorithm is implemented in 1-D, 2-D and 3-D and suitable examples will be presented.

The use of Logical operators as a new tool for Texture Classification problem is presented. Logical operators like Hadamard, Adding, Arithmetic, Conjunction, Disjunction and Equivalence are used to form a set of filters. The order-2 matrix of each operator is used for filtering the gray scaled textured image. Standard deviation matrices are computed from these filtered images over a 5 by 5 moving window. Features are extracted on this standard deviation matrix using zonal masks. Zonal masks with annular-ring, wedge, and parallel slit sampling geometries are used. Supervised classifier such as Euclidean is used in the experiment. Experiments are performed with Brodatz textures and also on remote sensing images. Nine different sets of 6 Bradatz textures are used. Two remote sensing images are classified using pixel to pixel classification. Out of the 33 Brodatz textures taken for classification, for 22 textures the Percentage of Correct Classification (PCC) is 100%, for 4 textures PCC is in the range of 95% to 99%. The algorithm developed is also computationally efficient as it involves only addition operations.

A concept to improve the accuracy in an unsupervised texture segmentation is presented in this paper. In a supervised segmentation, some known information or human intervention is provided to achieve a good segmentation result. However, unlike the supervised method, an unsupervised approach does not require any known information about the image to be segmented. This restriction can make the approach less accurate than the former method. To overcome such a restriction, we propose a mechanism to automatically obtain the information about the image before the actual segmentation process begins. This is possible by introducing a pre- segmentation step to obtain preliminary information about the image. Such self-generated information is called pseudo-a priori knowledge because it is not known or supplied to the segmentation process. The information is useful as a guideline in the selection of a critical set of Gabor filter parameters which in turn improve the accuracy in texture segmentation. The segmentation results of the proposed pseudo-supervised approach are compared with the results from a typical segmentation approach using a fixed quasi-complete Gabor filter set by applying them to the same test images. The segmentation results are compared using the percentage of misclustered pixels.

An important challenge in mapping image-processing techniques onto applications is the lack of quantitative performance measures. From a systems engineering perspective these are essential if system level requirements are to be decomposed into sub-system requirements which can be understood in terms of algorithm selection and performance optimization. Nowhere in computer vision is this more evident than in the area of image segmentation. This is a vigorous and innovative research activity, but even after nearly two decades of progress, it remains almost impossible to answer the question 'what would the performance of this segmentation algorithm be under these new conditions?' To begin to address this shortcoming, we have devised a well-principled metric for assessing the relative performance of two segmentation algorithms. This allows meaningful objective comparisons to be made between their outputs. It also estimates the absolute performance of an algorithm given ground truth. Our approach is an information theoretic one. In this paper, we describe the theory and motivation of our method, and present practical results obtained from a range of state of the art segmentation methods. We demonstrate that it is possible to measure the objective performance of these algorithms, and to use the information so gained to provide clues about how their performance might be improved.

This paper presents a general approach to fast image indexing and searching for content-based image retrieval on a network of workstation clusters. Three primary issues in image retrieval are discussed: image feature extraction and representation, similarity measure, and searching methods. A wavelet based image feature extraction scheme is introduced to represent images with multiple features such as colors, textures and shapes. In addition, a feature component code is proposed to facilitate a dynamic image indexing scheme where images are queried by different features or combinations. Furthermore, the relevance feedback technique for information retrieval is used to convert image feature vectors to weight- term vectors for efficient searching.

As more and more visual information is available on video, information indexing and retrieval of digital video data is becoming important. A digital video database embedded with visual information processing using image analysis and image understanding techniques such as automated target detection, classification, and identification can provide query results of higher quality. We address in this paper a robust digital video database system within which a target detection module is implemented and applied onto the keyframe images extracted by our digital library system. The tasks and application scenarios under consideration involve indexing video with information about detection and verification of artificial objects that exist in video scenes. Based on the scenario that the video sequences are acquired by an onboard camera mounted on Predator unmanned aircraft, we demonstrate how an incoming video stream is structured into different levels -- video program level, scene level, shot level, and object level, based on the analysis of video contents using global imagery information. We then consider that the keyframe representation is most appropriate for video processing and it holds the property that can be used as the input for our detection module. As a result, video processing becomes feasible in terms of decreased computational resources spent and increased confidence in the (detection) decisions reached. The architecture we proposed can respond to the query of whether artificial structures and suspected combat vehicles are detected. The architecture for ground detection takes advantage of the image understanding paradigm and it involves different methods to locate and identify the artificial object rather than nature background such as tree, grass, and cloud. Edge detection, morphological transformation, line and parallel line detection using Hough transform applied on key frame images at video shot level are introduced in our detection module. This function can also help rapidly filter incoming video and extract only those video sequences of potential interest under real time combating environment. Experimental results on video sequences acquired by Predator prove the feasibility of our approach.

Relaxation matching is one of the most relevant methods for image matching. The original relaxation matching technique using point patterns is sensitive to distortions such as missing or spurious points and random errors. In this paper, we present a robust point pattern relaxation matching technique based on least median of squared errors estimation to improve the original point pattern relaxation matching technique to be robust to missing or spurious points and random errors. Experimental results with large simulated images and real images demonstrate the effectiveness and feasibility of the method to perform point pattern relaxation matching with missing or spurious points and random errors.

This paper describes an approach to segmentation of textured grayscale images using a technique based on image filtering and the fractal dimension (FD). Twelve FD features are computed based on twelve filtered versions of the original image using directional Gabor filters. Features are computed in a window and mapped to the central pixel of this window. An iterative K-means-based algorithm which includes feature smoothing and takes into consideration the boundaries between textures is used to segment an image into a desired number of clusters. This approach is partially supervised since the number of clusters has to be predefined. The fractal features are compared to Gabor energy features and the iterative K- means algorithm is compared to the original K-means clustering approach. The performance of segmentation for noisy images is also studied.

An algorithm is proposed in order to estimate the rotation between two frames of a scene. Only linear segments and their geometric attributes are used and the algorithm is based on the correlation of the slope angle histograms of linear segments extracted from two frames. The basic idea is that from one frame to the other, if the camera rotates (theta) _r degrees, linear segments will also be rotated by (theta) _r degrees. In order to alleviate computational complexity a coarse to fine approach is proposed. The algorithm can be employed for image analysis tasks such as matching, registration, mosaicing images. Targeted applications are mainly in the domain of aerial images of man-made sites, and they cover the analysis of time-varying images. The algorithm is tested on both synthetic and real images. Synthetic frames are generated by a program which simulates, via its parameters similar effects as real images have.

Skeletonization as a tool for quantitative analysis of three- dimensional (3D) images is becoming more important, as they are more common in a number of application fields, especially in biomedical tomographic images at different scales. Here we propose a method, which computes both surface and curve skeletons of 3D binary images. The distance transform algorithm is applied to reduce a 3D object to a 2D surface skeleton, an then to a 1D curve skeleton in two phases. In surface skeletonization, 6-connectivity is used in distance transform; while in curve skeletonization, 18-connectivity is used in computing distance transform. Some examples are discussed to illustrate the algorithm.

Classification of earth terrain from satellite radar imagery represents an important and continually developing application of microwave remote sensing. The basic objective of this paper is to derive more information, through combining, than is present in any individual element of input data. Multispectral data has been used to provide complementary information so as to utilize a single SAR data for the purpose of land-cover classification. More recently neural networks have been applied to a number of image classification problems and have shown considerable success in exceeding the performance of conventional algorithms. In this work, a comparison study has been carried out between a conventional Maximum Likelihood (ML) classifier and a neural network (back-error-propagation) classifier in terms of classification accuracy. The results reveal that the combination of SAR and MSS data of the same scene produced better classification accuracy than either alone and the neural network classification has an edge over the conventional classification scheme.

In this paper the use of wavelets for classifying noisy textures is presented. The textures are decomposed using wavelets. The coefficients are thresholded based on a residual energy criterion. Restoration involves thresholding the wavelet coefficients only to a level at which the textures can be discriminated. The decomposition and thresholding is stopped when the energy based criterion is satisfied. A classifier is then used to discriminate the textures. Uniform, Gaussian, speckle and salt & pepper noise are added to the textures. Classification and segmentation experiments are conducted on photographic textures and remote sensing images. The algorithm performs well with all the types of noise, upto SNR's as low as 0 dB. The method is adaptive to any type of noise and gives improved performance compared to the available methods for texture discrimination in the presence of noise.

Many imaging systems utilize detector arrays that are not sufficiently dense to sample the scene according to the Nyquist criterion. As a result of this undersampling, some of the higher spatial frequencies admitted by the optics are aliased to lower frequency representations in the sampled image. This creates undesirable artifacts that decrease the utility of the imagery. Furthermore, the blurring effects of the optics and the finite detector size also degrade the image quality. Several approaches for obtaining a higher sampling rate have been suggested in the literature such as controlled and uncontrolled microscanning or from natural camera platform motion in order to up-sample the scene. Here we extend this work to include the possibility of non-global rigid body motion. In particular, we show that the motion of rigid objects within the scene is in many cases sufficient to allow the up-sampling of the object. We present a method that makes use of optical-flow-based motion segmentation techniques to isolate moving objects from the background in a sequence of digital images. These segmented objects from each frame provide a unique 'look' or set of samples of the object, allowing us to perform a reconstruction of the object. This allows for high resolution image reconstruction of each segmented object independently. The experimental results presented illustrate the breakdown of global reconstruction algorithms in the presence of non-global rigid motion. We also present results using the proposed method that treats individual moving objects and background separately. The results include data from a visible CCD camera.

Morphological filters are investigated and employed for detecting and visualizing objects within an image. The techniques developed here will be employed on NASA's Earth Observing System (EOS) satellite data products for the purpose of anomaly detection. Previous efforts have shown the phase information in the spectral domain to be more significant than the magnitude information in representing the location of objects in an image. The magnitude information does provide some useful information for object location, but it is also sensitive to image illumination, blurring, and magnification variations, all of which influence the performance of object detection algorithms. Magnitude reduction techniques in the spectral domain can dramatically improve subsequent object detection methods by causing them to rely less on the magnitude and more on the phase information of the image. We propose three new improvements to our object enhancement and detection techniques. Our first method is an enhancement to our previous magnitude-reduction technique. Our second improvement is a modification of our Rational Morphological Filters in which we raise our resulting image to a power, thereby magnifying our feature detection capability. Third, we look at speed enhancement by utilizing Hartley and Walsh Transforms in place of classical Fourier techniques.

Edge Enhancement is an important image processing method for remote sensing image data. Many images are compressed by JPEG standard which uses the Discrete Cosine Transform (DCT). Manipulating data in the DCT domain is an efficient way to save the computer resources. In this paper, new algorithms for edge enhancement of remote sensing image data in the DCT domain are developed and implemented in three steps: highpass filtering, adding back full or part of gray levels to the original image, and linear contrast stretching. In addition, a method to approximate the minimum (MIN) and maximum (MAX) gray level intensity is presented in the paper, which are the necessary information for contrast stretching. Experimental results show that the quality of images generated by the new algorithms are comparable with that of images generated by the corresponding methods in the spatial domain.

Subband image coding techniques have received considerable attention as a powerful source coding ones. These techniques provide good compression results, and also can be extended for progressive transmission and multiresolution analysis. In this paper, we propose a hybrid subband image coding scheme using DWT (discrete wavelet transform), DPCM (differential pulse code modulation), and ADPCM (adaptive DPCM). This scheme produces both simple, but significant, image compression and transmission coding.

Multidimensional phase space of a system attractor is transformed into the space constructed by distance topological sequences created from measured digitized signal. This transformation allows to reduce calculation complexity of attractor characteristics determination and experimental data quantity. It has been proved that representation of the attractor by means of distance topological sequences is sufficient for determination of all parameters of phase trajectories. Numerical simulations showed high convergence and reliability of proposed method.

A new method of restoration of images distorted by atmosphere is presented in this paper. In imaging through the atmosphere, the system MTF can be decomposed into the sum of its mean, <H>, and random component, (Delta) H. This technique is based on identification of Fried parameter r_o in expression of <H>. The identification of r_o is based on the facts: 1, Deterministic component, <H>, including average turbulence MTF and aerosol MTF, is located at low frequency areas in the frequency domain. 2, The shape of <H> can be analytically determined according to Fried theory. 3, <H> has the same zero-cross point as the image degraded by this <H>. A corrected r_o can result in a good restoration result. This is due to the main contribution on blur is <H>, while the random MTF, (Delta) H, can be removed by Wiener filter.

In this paper, we present a panorama system which extends the camera's field of view to the human's one. As the camera's field of view is much smaller than the human's, object could not be captured in a single frame in many cases. To solve this problem a fish-eye lens can be used. However the images obtained in this way have problems in image quality. Another solution is the panorama by frame-aligning and frame-pasting video frames in video sequence. Panorama System gets real time images from the panning video camera and takes align-paste process using registration algorithm. This System can show real time front scene in the human's field of view and can be used in the area-observation application. Experimental results prove the effectiveness and feasibility of the system.

A 3D terrain reconstruction method using compound techniques is proposed. Normal matching results only supply DSM (Digital Surface Model). This means matching results may be on the top of man-made objects such as houses or trees. This kind of matching results could not supply an accurate DEM (Digital Elevation Model). The proposed method is a more efficient method for determining elevations from overlapping digital aerial images and satellite images. It combines image analysis and image matching methods and supplies a more accurate DEM.

Building image mosaics is interesting to make a wide or a panoramic view from a set of small views, or to cover a large area with aerial images, for instance for cartography purposes or scene simulation. The images used in this experiment are overlapping enough to ensure a robust match, and the geometric transformation between images corresponds to an homographic transformation. As the mosaic will be interpreted, it must be geometrically reliable and must present no radiometric artifact. The homographic transformation between two overlapping images is initialized using the viewing parameters, or using a small set of corresponding points. To improve this first estimation, an original inter-image matching method is used. Radiometric correction is also done by comparing images on the overlap area. The mosaic is built by pasting images successively. Smoothing the intensity values near the seam-line is classically used to mask discontinuities, but that implies radiometric modifications which can lead to false interpretations. To solve this problem, especially if differences in the overlap area are still important after both corrections, a search for the most invisible seam-line is performed with a dynamic programming algorithm.

Image sharpening based on neural network (NN) approximation techniques is applied to increase the spatial resolution of Landsat thematic mapper (TM) thermal-infrared (T-IR) data. Sharpening is derived from a learned input-output mapping of image edge contrast patterns between T-IR and higher resolution reflective TM bands. This method is similar to a reported adaptive least squares (LS) method used to estimate TM T-IR data at a higher resolution. However, there are two major differences: use of NN approximation instead of LS estimation, and application of a reported multiresolution technique to combine spatial information adaptively from the original image and its high spatial resolution estimate. With training pair examples from reduced spatial resolution data, a multilayer feedforward NN is trained to approximate T-IR data samples from a small neighborhood of samples from three other TM bands. Output of the trained NN for full-resolution input data is an estimate of T-IR image at full resolution. One advantage of this method is that the NN approximator can be trained from a subset of image scene samples and yet be applied to the entire scene. Preliminary examples illustrate sharpening at four times higher resolution. The accuracy of the technique was evaluated with a simulated lower spatial resolution image that included blurring introduced by the TM sensor's PSF. Although results are promising, further evaluation with simulated lower resolution IR data is needed.

Developing computer technology provides a great opportunity to develop multimedia product over the Internet. An iterative Teleconference system is one of the most important applications, which has been widely used all over the world. Due to the characteristics of the Internet, it is difficult to guarantee the Quality of Service (QoS) when such a system is used on the Internet. Burstiness, synchronization among Audio and Video, and the reduced Internet bandwidth especially across national boundaries are the most important factors, which seriously may degrade the QoS. As an integrated solution for all these three problems, the SmartChannel was designed and implemented in this paper.

The microgravity environment of the NASA Shuttles and Russia's Mir space station have been measured by specially designed accelerometer systems. The need for comparisons between different missions, vehicles, conditions, etc. has been addressed by the two new processes described in this paper. The Principal Component Spectral Analysis (PCSA) and Quasi- steady Three-dimensional Histogram (QTH) techniques provide the means to describe the microgravity acceleration environment of a long time span of data on a single plot. As described in this paper, the PCSA and QTH techniques allow both the range and the median of the microgravity environment to be represented graphically on a single page. A variety of operating conditions may be made evident by using PCSA or QTH plots. The PCSA plot can help to distinguish between equipment operating full time or part time, as well as show the variability of the magnitude and/or frequency of an acceleration source. A QTH plot summarizes the magnitude and orientation of the low-frequency acceleration vector. This type of plot can show the microgravity effects of attitude, altitude, venting, etc.