This paper describes the effect of using lossy transform- based compression on IR framing sensor data with and without sensor anomaly pre-processing. Significant image degradation persists when ground-based non-uniformity correction processing is implemented after the sensor imagery has been compressed and reconstructed. Various techniques for non- uniformity correction, from low to high processing complexity, are applied before and after lossy compression at varying compression ratios. Results using both DCT and wavelet transform-based compression techniques indicate that on-board real-time compression algorithms and nonuniformity correction must be jointly optimized and that direct application of lossy compression without preprocessing for sensor anomalies reduces not only the compression efficiency and image fidelity, but also the performance of subsequent ground-based nonuniformity correction.

The near-lossless CALIC is one of the best near-lossless intraframe image coding schemes which exploits and removes the local context correlation of images. Wavelet transform localizes the frequency domain and exploits the frequency- based global correlation of images. Applying the context modeling for the wavelet transform coefficients, a state of the art intraframe near-lossless coding scheme can be obtained. In this paper, we generalize the intraframe wavelet transform CALIC to interframe coding to form a hybrid near-lossless multispectral image compression. Context modeling techniques lend themselves easily to modeling of image sequences. While wavelet transform exploits the global redundancies, the interframe context modeling can thoroughly exploit the statistical redundance is both between and within the frames. First, the image frame is wavelet transformed in the near-lossless mode to obtain a set of orthogonal subclasses of images. Then the coefficients of interframes of the image are predicted using the gradient-adjusted predictor based on both intra- and inter-frame current coefficient context. The predicted coefficients are adjusted predictor based on both intra- and inter-frame current coefficient context. The predicted coefficients are adjusted using the sample mean of prediction errors conditioned on the current context and the residues are quantized. An incremental scheme is used for the prediction errors in a moving time windows for prediction bias cancellation. All the components are distortion controlled in the minmax metric to ensure the near-lossless compression. The decompression is the inverse of the process. It is demonstrated that the near-lossless wavelet transform and context modeling interframe image compression is one of the best schemes in high-fidelity multispectral image compression and it outperforms its intraframe counterpart with 10-20 percent compression gains while keeping the high fidelity.

In this paper several adaptive order statistic filters (OSF) are developed and compared for channel characterization and noise suppression in images and 3D CT data. Emphasis has been put on the situation when a noise-free reference image is not available but instead we can have a sequence of two noisy versions of the same image. One of the noisy images is used as the reference in the OSF. It is shown theoretically that if noises are not correlated, the expected values of the derived filter coefficients will be equal to those coefficients derived using a noise-free reference. Experiments using the noisy reference images yield comparable result to those methods using a noise-free reference image nd also better results than those of median, Gaussian, averaging and Wiener filters.

In this paper, the fractal function is used for the image representation and compression. By introducing an searching strategy into the fractal-function-based image coding, the image representation accuracy is improved by 10 percent. A regional adaptive resolution-based fractal block coding is formed by combing the encoding and the decoding processes together with a quadtree structure to perform image compression. It is shown by the experiments that the image compression ratio is increased up to 12 without a great loss of visual quality.

We address multiresolutional encoding and decoding within the embedded zerotree wavelet (EZW) framework for both images and video. By varying a resolution parameter, one can obtain decoded images at different resolutions from one single encoded bitstream, which is already rate scalable for EZW coders. Similarly one can decode video sequences at different rates and different spatial and temporal resolutions from one bitstream. Furthermore, a layered bitstream can be generated with multiresolutional encoding, from which the higher resolution layers can be used to increase the spatial/temporal resolution of the images/video obtained from the low resolution layer. In other words, we have achieved full scalability in rate and partial scalability in space and time. This added spatial/temporal scalability is significant for emerging multimedia applications such as fast decoding, image/video database browsing, telemedicine, multipoint video conferencing, and distance learning.

The paper demonstrates the existence of a common autoregressive (CAR) compressing transform for a class of similar images, and examines the model sensitivity with respect to translations and rotations of the similar images.

The information-theoretic network for independent component analysis has been studied for unsupervised learning in the signal processing area. We derive a learning rule from the mutual information or the sum of the marginal entropy based on the local-Gaussian assumption for blind source separation of the convolutive mixture. The algorithm has been tested for several real-world recordings and showed the promising results.

Traditional fractal coding searches the best mapping domain block in a searching pool, applying 8 symmetry operation to each domain block to satisfy the tolerance condition, thus increasing coding complexity greatly. We here present a new fast fractal image compression method based on block variance and wavelet transform which increase the coding speed up to 25 times and the reconstructed image has no obvious degradation in visual quality.

The paper studies the use of common autoregressive models for a similar database compression and estimates compression ratio for predictive models based on correlation coefficient.

Electron backscatter patterns are used to analyze the crystallographic structure of materials and extract 3D grain orientation data that is of use in material science. We present new Hough transform pre- and post-processing algorithms to improve analysis of such data. Our new technique also reduces the computation time required.

The author shows that mathematical morphology imaging filter techniques enhance the effectiveness and versatility of maximum entropy thresholding in separating foreground and background, especially for small targets. Mathematical morphological image processing techniques, specifically openings and closing, tend to set large areas of a gray- level image to the same gray-level while preserving the number of gray-levels present in small areas, i.e., small targets. In an entropic analysis of the image, this equates to minimizing the entropy of the areas set to identical gray-levels, while conversely enhancing that of small, information-rich regions. Maximum entropy thresholding entropy contribution of each gray-level. Thus, prefiltering an image using an opening or closing operation immensely improves maximum entropy thresholding. Examples of this combined technique are shown for both one- and two- dimensional entropic thresholding. The author points to this synergism as an example of the inherent interconnectedness of image processing and thresholding algorithms, and emphasizes the importance of the analysis of combined algorithms in the design of target detection and tracking schemes.

In this paper a compact opto-electric unit (CHOEU) for digital image processing and recognition is proposed. The central part of CHOEU is an incoherent optical correlator, which is realized with a SHARP QA-1200 8.4 inch active matrix TFT liquid crystal display panel which is used as two real-time spatial light modulators for both the input image and reference template. CHOEU can do two main processing works. One is digital filtering; the other is object matching. Using CHOEU an edge-detection operator is realized to extract the edges from the input images. Then the reprocessed images are sent into the object recognition unit for identifying the important targets. A novel template- matching method is proposed for gray-tome image recognition. A positive and negative cycle-encoding method is introduced to realize the absolute difference measurement pixel- matching on a correlator structure simply. The system has god fault-tolerance ability for rotation distortion, Gaussian noise disturbance or information losing. The experiments are given at the end of this paper.

In hail flow analysis, hail characteristics such as isolated hail size histogram and hail velocity distribution are very important. An innovative automatic hail flow measuring system is described in this paper. The system applies Rewo filter based image processing, automatic learning, and multiprocessor techniques to analyze both temporal and spatial features of multiple hail flow video frames. The testing results of five hundred hail frames having a total of more than two thousand hail particles show the advantages of this system in automatic, robustness, statistic reliability and accuracy compared to radar and manual calculation.

We describe HOPS, a system for extracting some audiovisual cues for the modeling of a computer end-user environment. The objective of the study is to provide some reliable audiovisual cues in order to 'augment' the computer input devices set for multimodal applications. The system accepts an audio-visual scene as input and produces different kinds of events which could contribute to increase the awareness and robustness of interactive system. The described framework for the extraction of cues is ecological and homogenous. On the audio path a cross power spectrum method is applied for extracting different kind of acoustic patterns defined as acoustic segments. The acoustic signal from a microphone and the acoustic segments are firstly FFT- transformed, averaged, and secondly correlated in the spectral domain. The maxima of the inverse Fourier transform of this cross-power spectrum is the criteria for the detection of some acoustic events. On the video path, we define some initial color models of some desired cues such as mouth, eyes, etc. and then track them in the audiovisual scene recorded by a camera.

Over the last decade, various designs for hyperspectral instruments have been developed and may be categorized roughly by the way in which they acquire hyperspectral data: via filter, dispersion, or Fourier transform. Each category has unique characteristics that led to differing processing needs. Fueled by increasing demands for real time hyperspectral data from space and aircraft platforms, a new generation of data processing capabilities are being developed by an increasingly large community with the objective of accommodating the high data rate produced by these hyperspectral imagers. This paper provides an overview of the three basic categories of HSIs and then contrast each with respect to current and planned processing capabilities.

We describe the optical correlator subsystem of a hybrid vision system. The optical correlator has a Kopin SLM at the input plane and a fast analog nematic Boulder Nonlinear Systems SLM at the filter plane. Filters are computed to provide identification and pose estimation for various objects. Spatial non-uniformities in the filter SLM are particularly bothersome from an operational standpoint. Methods of measuring and accommodating for the non- uniformities are described. Correlation filters produce an initially coarse identification, and later filters are intended to refine the coarse estimate of the pose of the observed object. Certain image preprocessing operations and their results are discussed; these include the cross- correlation that result when a Laplacian of Gaussian image is used as input to the system whose filter has been computed for a Sobel edge-extracted version of the reference image.

We describe the digital subsystem of a hybrid vision system. The digital processing includes feature extraction, projection of an object's model through a camera model, space-domain correlation of the model feature with the observed feature map, and a constrained solution that minimizes errors in the fit between modeled and observed features by using a gradient search routine. The digital processing refines the coarse information from the optical processor. Examples will be shown of the degree to which the digital system can refine the precision of the optical estimate and how the optical pre-processing can provide a starting pose for the gradient search, thus minimizing the likelihood of the process's catching on a local minimum in the cost function.

This paper presents a ne method to calculate an image hologram pattern for a 3D object modeled as a stack of horizontally sliced 2D images. It is shown that the proposed method manipulates data on the angular spectrum domain, just as a recently presented method by Leseberg and thus it is regarded as an approximation to the Leseberg's method. Major differences arise in the implementation where it utilizes FFT in two perpendicular directions, which reduces the computational load considerably although additional interpolation process is required. We finally provide some comparisons between the proposed method and the existing methods in terms of computational complexities and performances.

Conventionally, edges are treated as either scalar or vector quantities. This paper presents a novel framework which treats edges as directional dipoles that induce the field around themselves. An analogy can be made between this concept and the interaction of magnetic dipoles with the magnetic field. The dipoles interact with the field and align themselves into a smooth contour configuration. This paper shows the effectiveness of the concept in edge linking and proposes efficient computational schemes for real-time implementation of the edge dipole interactions. It also proposes an image representation using the dipoles on a hexagonal lattice and a contour extraction algorithm implemented on the representation. The algorithm consists of three processes: noise removal, edge alignment and edge thinning/extension. The results of some experimental studies are also presented.

Classification of an image on the basis of contained patterns is considered in a context of detection and estimation theory. To simplify mathematical derivations, image and reference patterns are represented on a complex support. This allows to convert the four positional parameters into two complex numbers: complex displacement and complex scale factor. The latter one represents isotropic dilations with its magnitude, and rotations with its phase. In this context, evaluation of the likelihood function under additive Gaussian noise assumption allows to relate basic template matching strategy to wavelet theory. It is shown that using circular harmonic wavelets simplifies the problem from a computational viewpoint. A general purpose pattern detection/estimation scheme is introduced by decomposing the images on a orthogonal basis formed by complex Laguerre-Gauss Harmonic wavelets.

In this paper we consider the application of standard image processing algorithms to extract parameters from weather radar data that can be subsequently used in converting radar reflectivity Z to rainfall rate R. We also examine the possible advantage of using the total time rate of change of Z, Z, as a modification to the standard Z-R relationship. The rationale behind this approach is based on the observation that convective rainfall often produces drop size distributions (DSDs) which are significantly different from the DSDs produced by stratiform rainfall. This is somewhat similar in concept to previous rainfall regime classification strategy, using the spatial gradient of the horizontal reflectivity pattern to switch from a convective to a stratiform set of Z-R parameters. In this case, however, Z provides a continuous range of values rather than the binary values of the previous classification method. A time series of radar patterns are processed using standard image processing algorithms to segment the image into spatial areas corresponding to storm cells and multi-cell regions. These regions are tracked using thresholding and edge detection algorithms, as well as 2D cross-correlation, for determining the advection velocity. The average Z of each regions calculated and Z is then found by differences between consecutive frames. The final value of Z is used to modify the Z-R transformation for the corresponding region. Note that if Z is small, the resulting formula reduces to a standard form of the Z-R relationship. Similarly, the time rate of change of height at a constant dBZ level is considered as another approach to modifying the standard Z-R relation.

A method of the Euclidean distance map generation is proposed which reduces the number of multiplication operations used to compute distances. This method belongs to a class of the ordered propagation algorithms using masks whose shape depends on a direction of the distance value propagation. To obtain Euclidean distances, we apply two non-Euclidean transforms simultaneously so that our approach is faster than other techniques because it uses only additions instead of multiplication operations when labeling the distance map. Experiments confirm a correctness of our approach and memory requirements for it do not exceed those for other transforms with the ordered propagation.

We introduce novel robust filtering applicable to image processing. They were derived using RM-type point estimations and the restriction technique of the well-known specific for image processing KNN filter. The derived RM-KNN filters effectively remove impulsive noise while edge and fine details are preserved. The proposed filters were tested on simulated images and real data and were provided excellent visual quality of the processed images and good quantitative quality in the MSE sense over standard median filter. Recommendations to obtain best processing result by proper selection of derived filter parameters are given. Two derived filters are suitable for impulsive noise reduction in any image processing applications. One can use the RM-KNN filters as the first stage of image enhancement following by any non-robust techniques such as Sigma-filter on the second stage.

In this paper, we propose a new approach to deal with the translation-invariant problem of wavelet transform. We first adaptively renormalize the original signal by a orthonomal scale function and first order and second order moments of the signal. This procedure can be fulfilled by a changeable filter, of which the coefficients are selected from a 'mother mask' prepared before. Then, we decompose the renormalized signal using Mallat's algorithm, and prove that the coefficients of this adaptive wavelet transform, called adaptive wavelet invariant moments, are translation- invariant and scale-invariant. Finally, experiment result for 2D digital signals are given.