Byoung-Ho Kang

Device calibration of a color image scanner digitizing system by using neural networks

In the color image analysis, color images can be captured through color video cameras or color image scanners. The calibration of a color image capturing device is required as digitizing systems spectral sensibilities are not always identical to the human color matching functions. Calibration can be abstracted as a modeling problem from a device dependent color space to a device independent and uniform color space. In this paper, a neural network model is proposed for converting a device-dependent color space to a device-independent color space. Statistical, optimization, and genetic algorithm approaches are compared with the neural network method to find the best model with the smallest resultant /spl Delta/Ea*b*. Conclusions are drawn with experimental results.

Fuzzy inference and logical level methods for binary graphic/character image extraction

Thresholding is one of the most important approaches to image segmentation. It has been widely used to characterize many images containing some objects of reasonably uniform brightness against a background of differing brightness. Typical examples include handwritten/typewritten text and microscope bio-medical samples. Even though it can be applied to the image processing widely, there is no robust thresholding technique to circumvent noisy image. In this study, firstly, the published character/graphic image extraction techniques were reviewed and investigated and new thresholding technique such as fuzzy inference and modified logical level are proposed. In fuzzy inference technique, new methods of fuzzification, fuzzy rule, and defuzzification are introduced for lower error and high speed image binarization.

Learning rate updating schemes of unsupervised learning

This paper first presents the improved convergence method using adjust learning rate (/spl alpha/) and clustering analysis in Kohonens network. New learning rate method which adopts hyperexponential function to reduce the influence of order presented of input patterns improves the convergence time in training. We also propose the new error measuring method using Hamming distance between patterns of each class and cluster center, Secondly, new strategy to speed up training the GLVQ is also proposed. Instead of using learning update scheme by Pal et al. (1993), we introduce adaptive learning rate updating strategy. Though Pal et al.s approach needs a priori information, approach proposed works without external parameters, and convergence time is reduced compared with Pal et al.s.

Colour image difference analysis: quantitative versus qualitative methods

This paper describes investigation of statistical correlation between the results of two image colour difference criteria: qualitative and quantitative. The aim is that, provided there is consistently high correlation between the results of the different methods, the quantitative method, which is relatively more convenient, could be used instead of the much more laborious qualitative psychophysical method.

Adaptive hierarchical stereo matching using object segmentation and window warping

In this paper, we propose an adaptive stereo matching algorithm to encompass stereo matching problems in projective distortion region. Since the projective distortion region can not be estimated in terms of fixed-size block matching algorithm, we tried to use adaptive window warping method in hierarchical matching process to compensate the perspective distortions. In addition, probability theory was adopted to encompass uncertainty of disparity of points over the window in this study. The proposed stereo matching algorithm has been tested on both disparity map and 3D model view. The experimental results show that remarkable improvement is obtained in the projective distortion region.

A simultaneous gamut compression with extended line of cusps

In many industrial areas, exact color reproduction is needed as an aspect of the appearance of color on cross media. The paper introduces the new gamut mapping algorithm which uses the extended line of two cusps. The point in the original gamut has been mapped in the reproduction gamut in L*-C* space, and each coordinate is mapped simultaneously. The experiment showed that the proposed algorithm acquired a higher chroma (C*) than the previous algorithm and the use of the whole range of gamut.

A method of printer modeling with the spectral reflectance curves using error back propagation

Most color imaging devices exhibit color distortions due to nonlinear component characteristics. Among color correction techniques, the most popular is device characterization. Generally there are two parts in printer characterization using a neural network, i.e. printer modeling and printer controller. This paper presents a method of printer modeling with spectral reflectance values using error back propagation. In experiments, we used an ink jet printer. First we print many color patches using known CMY values, and measured these patches with spectral reflectance values at wavelengths. The data set of the measured values and the input values were split into the validation set for assessing the performance of the trained network and the training sets. For training we took CMY values for the input nodes and reflectance values for the target values of the output nodes. The input layer of the network receives input as a CMY vector. The output represents spectral reflectance at certain wavelengths. To get a better performance, we change the network scheme. For the evaluation, we transformed spectral reflectance values to XYZ values. With this model, we can make a printer controller, the reverse of a printer model. Printer controller takes XYZ values for input and CMY for output. If the printer prints with controller CMY, we have the same L*a*b* between printed color patches and input value of the printer controller. From a global point of view, this total system represents a mapping from a target L*a*b* value to the output L*a*b* value of the printer.

WYSIWYG colour generating system development

Three scanner characterisations and three different monitor characterisation methods were applied for developing WYSIWYG (What You See Is What You Get) system. Specially, the model developed by Berns et al. (1993, 1995) was evaluated under different sampling and optimisation methods, and compared with other characterisation techniques.

Development of a neural network building tool for pattern clustering/classification

A neural network building tool for pattern clustering/classification is developed. A conventional method, a fuzzy technique, and a neural network method are implemented. New schemes for fast convergence and better classification performances are developed for this tool.