It is time to reconsider our research priorities. The veterans in our community started out in an age when the image processing tools in the publishing industry were graphic arts cameras and screens, pin registers, knives, and goldenrod paper.

A set of multichannel camera systems and algorithms is described for recovering both the surface spectral-reflectance function and the illuminant spectral-power distribution from the data of spectral imaging. We describe a camera system with six spectral channels of fixed wavelength bands. This system is composed of a monochrome charge-coupled device camera, six different color filters, and a personal computer. The dynamic range of the camera is extended for sensing the high-intensity levels of highlights. We assume that the object surfaces in a scene are made of an inhomogeneous dielectric material whose reflection properties are described by the dichromatic reflection model. The process for estimating the spectral data comprises several steps: (1) finitedimensional linear model representation of wavelength functions, (2) illuminant estimation, (3) data normalization and image segmentation, and (4) reflectance estimation. An algorithm is proposed for detecting highlight areas in the image. The reliability of the camera system and the algorithms is demonstrated in an experiment. Finally, a new type of camera system using a liquid-crystal filter is proposed.

This paper describes a new type of multichannel color sensor with the special property of having all channels per pixel at the same spatial location. This arrangement is accomplished by stacking three amorphous thin film detectors on a substrate (glass or silicon), one on top of the other. It has the advantage that the color aliasing effect is avoided. This effect produces large color errors when objects of high spatial frequency are captured using multichannel imaging sensors with color filter arrays. The new technique enables the design of a three-channel sensor as well as a six-channel sensor, the latter being valuable even as a single-pixel device for metrology. In the six-channel case, color is captured in two ‘‘shots’’ by changing the bias voltages. The sensors are characterized colorimetrically, including methods such as multiple polynomial regression both for tristimulus and spectral reconstruction, and the smoothing inverse for spectral reconstruction. The results obtained with different types of regression polynomials, different sensors, and different characterization methods are compared. The results show that the new amorphous three-channel sensors perform comparably with commercial charge coupled device cameras under the conditions tested, while the six-channel’s performance is striking.

Before doing extensive color gamut experiments, we wanted to test the uniformity of CIE L*a*b*. This paper shows surprisingly large discrepancies between CIE L*a*b* and isotropic observation-based color spaces, such as Munsell: (1) L*a*b* chroma exaggerate yellows and underestimate blues. (2) The average discrepancy between L*a*b* and ideal is 27%. (3) Chips with identical L*a*b* hue angles are not the same color. L*a*b* introduces errors larger than many gamut mapping corrections. We have isotropic data in the Munsell Book. Computers allow threedimensional lookup tables to convert instantly any measured L*a*b* to interpolated Munsell Book values. We call this space ML, Ma, and Mb in honor of Munsell. LUTs have been developed for both LabtoMLab and MLabtoLab. With this zero-error, isotropic space we can return our attention to the original problem of colorgamut image processing.

Gamut mapping is a technique to transform out-of-gamut colors to the inside of the output device’s gamut. It is essential to develop effective mapping algorithms to realize ‘‘WYSIWYG’’ color reproduction. In this paper, three-dimensional gamut mapping using various color difference formulae and color spaces are considered. Visual experiments were performed to evaluate which combination of color difference formula and color space for gamut mapping was most preferred for five images. The color difference formulae used in the experiments were ?Eab * , ?Euv * , ?E94 , ?ECMC , ?EBFD , and ?Ewt . The color spaces used in the experiments were CIELAB, CIELUV, CIECAM97s, IPT, and NC-IIIC. A clipping method was used that maps all out-of-gamut colors to the surface of the gamut, and no change was made to colors inside the gamut. It was found that gamut mapping using ?E94 , ?ECMC , and ?Ewt were effective in CIELAB color space. For mapping images containing a large portion of blue colors, DEBFD and ?Euv * were found to be more effective. ?Eab * was least preferred for all images. With respect to color spaces, gamut mapping performed in the CIELUV color space was superior to any other color spaces for the blue region. We conclude that ?E94 in CIELUV and ?EBFD in CIELAB are the two most useful combinations of color difference formula and color space for gamut mapping, if we are to apply a single combination universally.

In color gamut mapping of pictorial images, the lightness rendition of the mapped images plays a major role in the quality of the final image. For color gamut mapping tasks, where the goal is to produce a match to the original scene, it is important to maintain the perceived lightness contrast of the original image. Typical lightnessremapping functions such as linear compression, soft compression, and hard clipping reduce the lightness contrast of the input image. Sigmoidal-remapping functions were utilized to overcome the natural loss in perceived lightness contrast that results when an image from a full dynamic range device is scaled into the limited dynamic range of a destination device. These functions were tuned to the particular lightness characteristics of the images used and the selected dynamic ranges. The sigmoidal-remapping functions were selected based on an empirical contrast enhancement model that was developed from the results of a psychophysical adjustment experiment. The results of this study showed that it was possible to maintain the perceived lightness contrast of the images by using sigmoidal contrast enhancement functions to selectively rescale images from a source device with a full dynamic range into a destination device with a limited dynamic range.

Solving the color constancy problem in many applications implies the understanding of chromatic adaptation. The Retinex theory justifies chromatic adaptation, as well as other color illusions, on visual perception principles. Based on the above theory, we have derived an algorithm to solve the color constancy problem and to simulate chromatic adaptation. The evaluation of the results depends on the kind of applications considered. Since our purpose is to contribute to the problem of color rendering for photorealistic image synthesis, we have devised a specific test approach. A virtual ‘‘Mondrian’’ patchwork has been created by applying a rendering algorithm with a photorealistic light model to generate images under different light sources. Trichromatic values of the computer generated patches are the input data for the Retinex algorithm, computing new color corrected patches. The Euclidean and the ?E94 * distances in the CIELAB space, between the original and Retinex color corrected trichromatic values, have been calculated. A preliminary analysis of the just noticeable difference has also been done on some colors compared to the closest MacAdam ellipses. Our work shows that the Retinex computational model is very well suited to solve the color constancy problem without any a priori information on the illuminant spectral distribution.

New applications such as printing on demand and personalized printing have increased the need for efficient lossless halftone image compression algorithms to lower the transmission time and the storage costs. State-of-the-art lossless bilevel image compression schemes like JBIG achieve only moderate compression ratios because they do not fully take into account the special image characteristics. In this paper, we present an improvement on the context modeling scheme by adapting the context template to the special patterns of halftone images. This is a nontrivial problem for which we propose a fast and efficient context template selection scheme based on the sorted autocorrelation function of a part of the image. We have experimented with classical halftones of different resolutions and sizes, screened under different angles, as well as with stochastic halftones. For classical halftones, the global improvement with respect to JBIG in its best mode is about 30%–50%. For stochastic halftones, the autocorrelation-based template gives no improvement, though a much slower exhaustive search technique shows that gains up to 70% are feasible using a suboptimal template. Binary tree modeling increases the compression ratio by another 5%–10%. Context modeling can also be used for other types of halftone image processing.

In the prepress industry, color images have both a high spatial and a high color resolution. Such images require a considerable amount of storage space and impose long transmission times. Data compression is desired to reduce these storage and transmission problems. Because of the high quality requirements in the prepress industry, mostly only lossless compression is acceptable. Most existing lossless compression schemes operate on gray-scale images. In this case the color components of color images are compressed independently. However, higher compression ratios can be achieved by exploiting intercolor redundancies. In this paper we present a comparison of three state-of-the-art lossless compression techniques which exploit such color redundancies: intercolor error prediction and a Karhunen–Loe`ve transform-based technique, which are both linear color decorrelation techniques, and interframe CALIC, which uses a nonlinear approach to color decorrelation. It is shown that these techniques are able to exploit color redundancies and that color decorrelation can be done effectively and efficiently. The linear color decorrelators provide a considerable coding gain (about 2 bpp) on some typical prepress images. Surprisingly, the nonlinear interframe CALIC predictor does not yield better results.

As we approach the new millenium, error diffusion is approaching the 25th anniversary of its invention. Because of its exceptionally high image quality, it continues to be a popular choice among digital halftoning algorithms. Over the last 24 years, many attempts have been made to modify and improve the algorithm—to eliminate unwanted textures and to extend it to printing media and color. Some of these modifications have been very successful and are in use today. This paper will review the history of the algorithm and its modifications. Three watershed events in the development of error diffusion will be described, together with the lessons learned along the way.

Color error diffusion can be classified into two types, namely: vector error diffusion and scalar error diffusion, according to the underlying quantization methods. Compared to scalar error diffusion, vector error diffusion is typically superior in overall image quality. However, it requires significantly more computation. In addition, it may occasionally introduce boundary artifacts. A color band, at a width of a few pixels to tens of pixels, may appear along the edges. In this paper, we propose two improvements in quantization for color error diffusion. First, we introduce an algorithm, which we call semivector quantization, to reduce the computational complexity of vector error diffusion. Second, we present a simple yet effective method for reducing boundary artifacts.

A new technique for building stochastic clustered-dot screens is being proposed. A large dither matrix comprising thousands of stochastically laid out screen dots is constructed by first laying out the screen dot centers. Screen dot centers are obtained by placing discrete disks of a chosen radius at free cell locations when traversing the dither array cells according to either a discretely rotated Hilbert space-filling curve or a random space-filling curve. After Delauney triangulation of the screen dot centers, the maximal surface of each screen dot is computed and isointensity regions are created. This isointensity map is converted into an antialiased gray scale image, i.e., into an array of preliminary threshold values. These threshold values are renumbered to obtain the threshold values of the final dither threshold array. By changing the disk radius, the screen dot size can be adapted to the characteristics of particular printing devices. Larger screen dots may improve the tone reproduction of printers having important dot gain.

The problem of shrinking the connected components of binary images to small residues is addressed. A new parallel algorithm is presented with the lowest known worst case time complexity as measured by iteration counts, i.e., ?(nc+nr)/2? iterations where the foreground of the image is contained within a rectangular region of size ncxnr pixels. The new algorithm also demonstrates superior average iteration counts in comparison with other shrinking algorithms on test images. The new algorithm uses a fully parallel application of an operator with an 8 pixel support. It reduces all connected components in any two-dimensional (2D) binary image to certain small 1, 2, or 3 pixel residues. It is potentially useful for component counting, as a fundamental step in connected component labeling algorithms implemented with fine grain realizations in 2D mesh computing architectures, and in other applications requiring preservation of foreground topology (connectivity).

This paper presents an aliasing reduction method for digital still cameras (DSCs) which takes advantage of the colorsignal correlation. The authors have developed a method of reducing aliasing by increasing the resolution of the color signals in DSCs using a single-chip charge-coupled device with a primary-color filter array. The new method interpolates the signals by taking advantage of the fact that color-signal variations are similar to one another in a local region of an image. In addition, it selects the signal that shows the highest correlation, either in the horizontal or vertical direction, for the color-signal positions subject to interpolation, thus improving color resolution. The effectiveness of the new method has been established: tests show that it significantly reduces aliasing.

Novel robust filtering algorithms applicable to signal and image processing are introduced. They use the robust techniques of rank and M estimators. The proposed robust rank M filters enhance the quality of radar images providing effective speckle suppression, impulse noise removal, preservation of edges and details. The properties of these algorithms are demonstrated by simulated images and real data. It is shown that the proposed filters provide good visual quality of processed images and have some advantages (in the MSE sense) in comparison to the known ones when the noise level is high.

Wavelet theory has been one of the most exciting developments in the past decade that brought together researchers from several different fields such as engineering, physics, and mathematics.