This paper presents a new approach for the design of feature-extracting recognition networks that do not require expert
knowledge in the application domain. Feature-Extracting Recognition Networks (FERNs) are composed of
interconnected functional nodes (feurons), which serve as feature extractors, and are followed by a subnetwork of
traditional neural nodes (neurons) that act as classifiers. A concurrent evolutionary process (CEP) is used to search the
space of feature extractors and neural networks in order to obtain an optimal recognition network that simultaneously
performs feature extraction and recognition. By constraining the hill-climbing search functionality of the CEP on specific
parts of the solution space, i.e., individually limiting the evolution of feature extractors and neural networks, it was
demonstrated that concurrent evolution is a necessary component of the system. Application of this approach to a
handwritten digit recognition task illustrates that the proposed methodology is capable of producing recognition
networks that perform in-line with other methods without the need for expert knowledge in image processing.
Linear Pixel Shuffling (LPS) dithering produces blue-noise-like patterns, but the placement of thresholds in a dither matrix is a result of an exact algebra, rather than iterative procedure -- as is usually the case. In this paper, we investigate the potential use of LPS for construction of color (CMYK) dithering masks.
In case of LPS dithering, the addition of the same value to each mask threshold, using modular arithmetic, is equivalent to the spatial mask shift. We propose a set of three shifted color masks for C, M, and Y that we construct from the original LPS mask using modular arithmetic. The main advantage of this approach is its simplicity. These shifts can be "tailored" to the statistical properties of the image and the set of new screens can be calculated on the fly.
The proposed method enables creation of screens of arbitrary size, since the dithering masks are tiled automatically (actually, the masks are of unlimited size). The number of gray levels in each screen is limited by the choice of a modulus number used for mask thresholds calculation. This enables us to use a virtually unlimited number of thresholds that are not necessarily linearly related to the LPS calculated matrix values. Thus, it is relatively easy to construct a non-linear dither screen that will compensate for any printer non-linearity.
We present a technique for converting continuous gray-scale images to halftone (black and white) images that lend themselves to lossless data compression with compression factor of three or better.
Our method involves using novel halftone mask structures which consist of non-repeated threshold values. We have versions of both dispersed-dot and clustered-dot masks, which produce acceptable images for a variety of printers.
Using the masks as a sort key allows us to reversibly rearrange the image pixels and partition them into groups with a highly skewed distribution allowing Huffman compression coding techniques to be applied. This gives compression ratios in the range 3:1 to 10:1.
A model to predict colorimetric value for color printers is presented. The Neugebauer narrow-band color mixing model was applied with modifications. While sixteen primaries are used for four-color printing process in Neugebauer mode, we used two data sets in our model, one with eighty-one CMYK primaries and the other with one hundred twenty-five CMY primaries. Two Yule-Nielsen factors were applied to optimize the CMYK set and the CMY set separately. The Yule-Nielsen factors were optimized by minimizing (Delta) E*<SUB>L*a*b*</SUB> or (Delta) E*<SUB>94</SUB>. The Neugebauer calorimetric quality factor was applied as a weighting function to optimize dot areas. By optimizing primaries and applying the CQF weighting function, the average color error and the maximum color error decrease significantly.
We investigate a method of ordering pixels (the elements of a rectangular matrix) based on an arithmetic progression with wrap-around (modular arithmetic). For appropriate choices of the progression's parameters, based on a generalization of Fibonacci numbers and the golden mean, we find equidistributed collections ofpixels formed by subintervals of the pixelprogression or "shuffle."
We illustrate this equidistributivity with a novel approach to progressive rendering of a synthetic image, and we suggest several opportunities for its application to other areas of image processing.
In the context of colorimetric matching, the intent of color scanner and printer calibrations is to characterize the devicedependent responses to the device-independent representations such as CIEXYZ or CIE 1976 L*a*b* (CIELAB). Usually, this is accomplished by a two-step process of gray balancing and a matrix transformation, using a transfer matrix obtained from multiple polynomial regression. Color calibrations, printer calibrations in particular,
are highly nonlinear. Thus, a new technique, the neural network with the Cascade Correlation learning architecture, is employed for representing the map of device values to CIE standards. Neural networks are known for their capabilities to learn highly nonlinear
relationships from presented examples. Excellent results are obtamed using this particular neural net; in most training sets, the average color differences are about one Eab. This approach is compared to the polynomial approximations ranging from a 3-term
linear fit to a 14-term cubic equation. The results from training sets indicate that the neural net outperforms the polynomial approximation. However, the comparison is not made in the same ground
and the generalizations, using the trained neural net to predict relationships it has not been trained with, are sometimes rather poor. Nevertheless, the neural network is a very promising tool for use in
color calibrations and other color technologies in general.