Heui-keun Choh

Image-based photorealistic 3-D face modeling

In this paper, we describe an automatic system for 3-D photorealistic face modeling from frontal and profile images taken by an uncalibrated handheld digital camera. The system employs a generic model adaptation framework. That is, after the fiducial features are detected from both images, a generic head model is deformed to match with the detected features. Realistic texture is created by combining the augmented facial textures from input images and synthesized texture, and mapped onto the deformed generic head model. Our shape deformation and texture generation algorithms have several advantages: (1) they generate photorealistic models even when the orthogonal assumption of input image pairs is not satisfied; (2) the generated ear has very accurate smooth shape and improves the quality of the whole face model dramatically; (3) a few supportive techniques on texture processing are employed to improve the visual quality.

Luminance contrast and chromaticity contrast preference on the colour display for young and elderly users

Abstract The human visual system changes with aging and one of the most important changes is the decrease of spatial contrast sensitivity. We investigated this change both for luminance contrast and chromaticity contrast, and both for threshold contrast and preferred contrast, (preferred by users to carry out a visual recognition task), in a series of psycho-physical experiments with achromatic and chromatic sinusoid gratings of different values of spatial frequency, hue, and luminance level, and with two observer groups: young and elderly observers. We investigated the spatial frequency range of 0.1–10 cycles per degrees. Our results indicate that, beyond the expected luminance contrast sensitivity decline of the elderly observers, the difference between the preferred luminance contrast of the elderly and the preferred luminance contrast of the young is even more significant than the threshold difference. The small preference differences between the age groups for chromaticity contrast compared to luminance contrast suggests that while with increasing age both the chromatic and the achromatic contrast sensitivity drops, preferred contrast stays more stable for chromaticity contrast than for luminance contrast.

Illuminant direction and shape of a bump

An algorithm for recovering illuminant direction from the image data of a smooth Lambertian surface illuminated with a distant pointwise light source is presented. The algorithm is based on analysis of intensity distributions around structural elements of the image, such as image regions corresponding to bumps. After recognition of a bump in the image and estimation of the illuminant direction, the image data are integrated in order to recover the shape of the surface patch. The shape integration algorithm, which is based on a novel theoretical approach, is used to compute the normal vector field in the bump region without use of any explicitly given initial curve or initial data about the bump region. The theoretical consideration is illustrated with some results achieved on simulated and on realistic images.

Subpixel arrangements and color image rendering methods for multiprimary displays

Design principles are formulated to develop visually optimal multiprimary subpixel architectures. Two new hexagonal subpixel architectures optimized for multiprimary color displays are shown, each designed according to these principles. These new multiprimary pixel architectures are considered to be useful to eliminate the color fringe artifact. They are expected to yield images of better visual quality than previous three primary color architectures including the RGB stripe architecture. A new image rendering method is also formulated for multiprimary subpixel architectures. This method can be used with and without subpixel rendering. An error function is defined enabling proper chromaticity reproduction and enhanced luminance resolution.

Gradient approach to quantify the gradation smoothness for output media

We aim to quantify the perception of color gradation smoothness using objectively measurable properties. We propose a model to compute the smoothness of hardcopy color-to-color gradations. It is a gradient-based method that can be determined as a function of the 95th percentile of second derivative for the tone-jump estimator and the fifth percentile of first derivative for the tone-clipping estimator. Performance of the model and a previously suggested method were psychophysically appreciated, and their prediction accuracies were compared to each other. Our model showed a stronger Pearson correlation to the corresponding visual data, and the magnitude of the Pearson correlation reached up to 0.87. Its statistical significance was verified through analysis of variance. Color variations of the representative memory colors-blue sky, green grass and Caucasian skin-were rendered as gradational scales and utilized as the test stimuli.

Perceptual gamut mapping on the basis of image quality and preference factors

This study has three primary aims circumventing current limitations of color reproduction technologies: firstly, to derive base-line image quality factors from both color printer experts and academic research works. Individual factors were verified by systematic experiments, secondly, to develop a perceptual gamut mapping algorithm covering the image quality and preference factors derived, thirdly, to apply the algorithm to printer driver as acting for a vendor specific perceptual intent. Algorithm of this study tried to optimization between control parameters of gamut mapping and color shifting factors of preference, e.g. skin, sky and green grass. Profile builder using this algorithm outperforms, in industrial and academic aspects, existing commercial tool and CIE recommended algorithms.

Black extraction method using gamut boundary descriptors

Color data conversion between CMYK and CIEL*a*b* color space is not directly corresponded, that is many CMYK combinations could reproduce the same CIEL*a*b* value. When building a LUT converting from CIEL*a*b* to CMYK for a CMYK color printer, one to one correspondence between CMYK and CIEL*a*b* must be aimed. The proposed method in this paper follows steps: (1) print and measure CIEL*a*b* values of CMYK reference chart, (2) set-up parameters to assign the amount of black extraction, (3) generate gamut boundary descriptors for gamut mapping and for black extraction using CMYK-CIEL*a*b* data under predetermined black extraction parameters, (4) perform gamut mapping for given CIEL*a*b* using the gamut boundary descriptor for gamut mapping, (5) determine K value of the gamut-mapped CIEL*a*b* using the gamut boundary descriptors for black extraction. The suggested method determines K value for given CIEL*a*b* using gamut boundary descriptors in CIEL*a*b color space. As a result, a color printer using this method can make out accurate black amount and reproduces more consistent CMYK images under different black extraction options.

Quantification of perceived macro-uniformity

Macro-uniformity refers to the subjective impression of overall uniformity in the print sample. By the efforts of INCITS W1.1 team, macro-uniformity is categorized into five types of attributes: banding, streaks, mottle, gradients, and moiré patterns, and the ruler samples are generated with perceptual scales. W1.1 macro-uniformity ruler is useful for judging the levels of print defect, but it is not an easy task to reproduce the samples having the same perceptual scales at different times in different places. An objective quantification method is more helpful and convenient for developers to analyze print quality and design printing system components. In this paper, we propose a method for measuring perceived macro-uniformity for a given print using a flat-bed scanner. First, banding, 2D noise, and gradients are separately measured, and they are converted to the perceptual scales based on subjective results of each attribute. The correlation coefficients between the measured values of the attributes and the perceptual scales are 0.92, 0.97, and 0.86, respectively. Another subjective test is performed to find the relationship between the overall macro-uniformity and the three attributes. The weighting factors are obtained by the experimental result, and the final macro-uniformity grade is determined by the weighted sums of each attribute.

New measurement method of banding using spatial features for laser printers

The techniques of one-dimensional projection in the spatial domain and contrast sensitivity function (CSF) are generally used to measure banding. Due to the complex printing process of laser printers, hardcopy prints contain other 2D nonuniformities such as graininess and mottle besides banding. The method of 1D projection is useful for extracting banding, but it induces the confounding effect of graininess or mottle on the measurement of perceived banding. The appearance of banding in laser printers is more similar to the sum of various rectangular signals having different amplitudes and frequencies. However, in many cases banding is modeled as a simple sinusoidal signal and the CSF is frequently applied. In this paper, we propose new measurement method of banding well correlated with human perception. Two kinds of spatial features give a good performance to banding measurement. First the correlation factor between two adjacent 1D signals is considered to obtain banding power which reduces the confounding effect of graininess and mottle. Secondly, a spatial smoothing filter is designed and applied to reduce the less perceptible low frequency components instead of using the CSF. By using moving window and subtracting the local mean values, the imperceptible low frequency components are removed while the perceptible low frequency components like the sharp edge of rectangular waves are preserved. To validate the proposed method, psychophysical tests are performed. The results show that the correlations between the proposed method and the perceived scales are 0.96, 0.90, and 0.95 for black, cyan, and magenta, respectively.

Hardcopy global color correction

When time, temperature or an external environment change, a laser electrophotographic printer produces quite different color tones from original ones. To achieve consistent color reproduction, many researchers have tried to characterize printer tone curves and developed methods to correct color tones. Color channel independent methods are most widely used, and there are two approaches in color channel independent method: (1) Instrument-based correction and (2) visual correction. Two approaches provide some trade-offs between cost and accuracy. In this paper we propose a methodology which combines the strengths of these two approaches. We describe how we design a calibration page and how we characterize lightness variation of a reference patch. We then present the procedure of our global tone correction method based on visual appearance match of end-users as well as the predetermined reference lightness model. We simulate tone distortion state by varying hardware parameters, and perform visual appearance match experiments to subjects. Our experimental results show that our method can significantly reduce color difference between the original print and the print at the distortion state. This suggests that we can reliably estimate the distortion parameter, and correct tones close to an original state.