The idea of this special section arose from the NMBIA98 workshop, organized by Steve Marshall in Glasgow in July 1998 as a spin-off of the European ESPRIT long-term research project, Nonlinear Model-Based Analysis and Description of Images for Multimedia Applications—Noblesse.

Filter design involves a trade-off between the size of the filter class over which optimization is to be performed and the size of the training sample. As the number of parameters determining the filter class grows, so too does the size of the training sample required to obtain a given degree of precision when estimating the optimal filter from the sample data. A common way to moderate the estimation problem is to use a constrained filter requiring less parameters, but then a trade-off between the theoretical filter performance and the estimation precision arises. The overall result strongly depends on the constraint type. Approaches presented in this paper divide the filter operation into two stages and apply constraints only to the first stage. Such filters are advantageous since they are fully optimal with respect to certain subsets of the filter window. Error expression, representation, and design methodology are discussed. A generic optimization algorithm for such two-stage filters is proposed. Special attention is paid to three particular cases, for which properties, design algorithms, and experimental results are provided: two-stage filters with linearly separable preprocessing, two-stage filters with restricted window preprocessing, and twostage iterative filters.

A new algorithm for removing mixed noise from images based on combining an impulse removal operation with local adaptive filtering in transform domain is proposed in this paper. The key point is that the operation is designed so that it removes impulses while maintaining as much as possible of the frequency content of the original image. The second stage is an adaptive denoising operation based on local transform. The proposed algorithm works well in denoising images corrupted by a white (Gaussian, Laplacian, exponential) noise, impulsive noise, and their mixtures. Comparison of the new algorithm with known techniques for removing mixed noise from images shows the advantages of the new approach, both quantitatively and visually. In this paper we also apply transformbased denoising methods for removing blocking and ringing artifacts from decompressed block transform or wavelet coded images. The method is universal and applies to any compression method used.

Grain noise is one of the most common distortions in cinematographic film sequences and is caused by the crystal structure of the chemical coating of the film material. The color-sensitive crystals can be considered as three separate populations. Thus, noise in the three channels is uncorrelated and similarly noise between frames is uncorrelated. Conversely, the signal (i.e., the projected view volume) is highly correlated between channels and over time. We shall explore methods of using this constraint to reduce noise within an adaptive filter framework using the popular Widrow– Hopf least-mean-square algorithm. As a film sequence typically includes many moving elements, such as actors on a moving background, motion estimation techniques will be used to eliminate as much as possible the effect of gray-level variations on the adaptive filter. An optical-flow technique is used to extract pixel motions prior to the application of the noise reduction.

This paper describes a method which uses the skull as a landmark for automatic registration of computer tomography to magnetic resonance (MR) images. First, the skull is extracted from both images using a new creaseness operator. Then, the resulting creaseness images are used to build a hierarchic structure which permits a robust and fast search. We have justified experimentally the performance of several choices of our algorithm, and we have thoroughly tested its accuracy and robustness against the wellknown mutual information method for five different pairs of images. We have found both comparable, and for certain MR images the proposed method achieves better performance.

Most verbal communications use cues from both the visual and acoustic modalities to convey messages. During the production of speech, the visible information provided by the external articulatory organs can influence the understanding of the language, by interpreting the combined information into meaningful linguistic expressions. The task of integrating speech and image data to emulate the bimodal human interaction system can be depicted by developing automated systems. These systems have a wide range of applications such as videophone systems, where the interdependencies between image and speech signals can be exploited for data compression and in solving the task of lip synchronization which has been a major problem. Therefore the objective of this work is to investigate and quantify this relationship such that the knowledge gained will assist in longer term multimedia and videophone research.

Nonlinear diffusion processes and watershed algorithms have been well studied for gray-scale image segmentation. In this paper we extend the use of these techniques to color or multichannel images. First, we formulate a general definition for a nonlinear diffusion process using the concept of an activity image that can be calculated for several image components. Then, we explain how the final activity image, obtained as a result of the nonlinear diffusion process, is fed through a watershed algorithm, yielding the segmentation of the image. The qualitative performance of the algorithm is illustrated with results for both gray-scale and color photographic images. Finally, we discuss the segmentation results obtained using a few well-known color spaces and demonstrate that a color principal component analysis gives the best results.

In this paper operations based on mathematical morphology which have been developed for binary and grayscale images are extended to color images. We investigate two approaches for ‘‘color morphology’’—a vector approach and a component-wise approach. New vector morphological filtering operations are defined, and a set-theoretic analysis of these vector operations is presented. We also present experimental results comparing the performance of the vector approach and the component-wise approach for multiscale color image analysis and for noise suppression in color images.

Histogram equalization and specification have been widely used to enhance information in a gray scale image, with histogram specification having the advantage of allowing the output histogram to be specified as compared to histogram equalization, which attempts to produce an output histogram that is uniform. Unfortunately, expanding histogram techniques to color images is not very straightforward. Since humans are sensitive to chromatic changes, care must be taken to ensure that incorrect colors are not produced. Additionally, expanding the one-dimensional histogram used in gray level histogram techniques to a joint histogram (usually of three variables representing the primary colors of red, green, and blue) can yield specified histograms which have no physical meaning, hence making it difficult to determine the set of histograms required for a desired enhancement. In this paper, we describe a method of extending gray level histogram specification to color images by performing histogram specification on the luminance, saturation, and hue components in the color difference C-Y color space. These methods take into account the relationship between the luminance and saturation components while yielding specified histograms that produce natural-looking results.

Some key issues related to space-based compression design are discussed. Various system considerations as well as potential compression options are also presented. A brief overview of a previously reported robust lossy transform coding algorithm is given followed by a study of its performance sensitivities. These sensitivities include (1) performance sensitivity to commonly observed anomalies in the data including band misalignment and dead/ saturated pixels, (2) impact of geometric distortion (processed versus unprocessed data) on compression performance, (3) performance sensitivity to different grouping of bands for spectral decorrelation, and (4) impact of compression on spectral fidelity. In addition, the impact of compression on the results of exploitation of environmental data including automated cloud study will be considered. It is shown that preprocessing to correct any geometric distortion noticeably improves the compression performance. Different groupings of bands also influence the performance. The loss of spectral fidelity, measured by the deviation from the original correlation coefficient matrix, is very insignificant regardless of the image and the coding bit rates. For the available bit rate, it is possible to trade off the compression-induced error between the spectral and spatial resolutions. In the implementation scenarios investigated, it was found that compression at rates approaching 16:1 has a minor impact on the exploitation and assessment of the ultimately derived automated cloud analysis. Additional work is needed to evaluate the impact of compression on other products, such as sea surface temperature. The results to date suggest that lossy compression may play a role in the efficient transmission of environmental information and in its subsequent exploitation.

Multilevel halftoning (multitoning) is an extension of bitonal halftoning, in which the appearance of intermediate tones is created by the spatial modulation of more than two tones, i.e., black, white, and one or more shades of gray. In this paper, the conventional multitoning approach and a previously proposed approach, both using stochastic screen dithering, are investigated. A human visual model is employed to measure the perceived halftone error for both algorithms. The performance of each algorithm at gray levels near the printer’s intermediate output levels is compared. Based on this study, a new overmodulation algorithm is proposed. The multitone output is mean preserving with respect to the input and the new algorithm requires little additional computation. It will be shown that, with this simple overmodulation scheme, we will be able to manipulate the dot patterns around the intermediate output levels to achieve desired halftone patterns. Implementation issues related to optimal output level selection and inkjet-printing simulation for this new scheme will also be reported.

Since the world is usually considered a three-dimensional environment, it would seem natural for image processing to attempt to recover and utilize as much information about the three-dimensional aspects of an image as possible. That has not been the case, mainly because the process of recovering three-dimensional information in a reliable and simple manner has been elusive.