We report on subjective experiments comparing example-based regularization, total variation regularization,
and the joint use of both regularizers. We focus on the noisy deblurring problem, which generalizes image
superresolution and denoising. Controlled subjective experiments suggest that joint example-based regularization
and total variation regularization can provide subjective gains over total regularization alone, particularly when
the example images contain similar structural elements as the test image. We also investigate whether the
regularization parameters can be trained by cross-validation, and we compare the reconstructions using crossvalidation
judgments made by humans or by fully automatic image quality metrics. Experiments showed that of
five image quality metrics tested, the structural similarity index (SSIM) correlates best with human judgement
of image quality, and can be profitably used to cross-validate regularization parameters. However, there is a
significant quality gap between images restored using human or automatic parameter cross-validation.
We investigate a Wiener fusion method to optimally combine multiple estimates
for the problem of image deblurring given a known blur and a corpus of sharper training images.
Nearest-neighbor estimation of high frequency information from training images is fused
with a standard Wiener deconvolution estimate. Results show an improvement in sharpness
and decreased artifacts compared to either the standard Wiener filter or the nearest-neighbor
reconstruction.
Custom color transformations for images or video can be learned from a small set of sample color pairs by estimating a look-up table (LUT) to describe the enhancement and storing the LUT in an International Color Consortium profile, which is a standard tool for color management. Estimating an accurate LUT from a small set of sample color pairs is challenging. Local linear and ridge regression are tested on six definitions of neighborhoods for twenty color enhancements and twenty-five color images. Excellent results were obtained with local ridge regression over proposed enclosing neighborhoods, including a variant of Sibson's natural neighbors. The evaluation of the different estimation methods for this task compared the fidelity of the learned color enhancement to the original sample color pairs and the presence of objectionable artifacts in enhanced images. These metrics show that enclosing neighborhoods are promising adaptive neighborhood definitions for local classification and regression.
A spatially-adaptive method for color printing is proposed that is robust to printer instabilities, reproduces
smooth regions with the quality of ordered dither, reproduces sharp edges significantly better than ordered
dither, and may be less susceptible to moire. The new method acts in parallel on square, non-overlapping blocks
of each color plane of the image. For blocks with low spatial activity, standard ordered dither is used, which
ensures that smooth regions are printed with acceptable quality. Blocks with high spatial activity are halftoned
with a proposed variant of dither, called ranked dither. Ranked dither uses the the same ordered dither matrix
as standard dither, but the ranks of the thresholds are used rather than the thresholds themselves. Ranked
dither is more sensitive than ordered dither to edges and more accurately reproduces sharp edges. Experiments
were done with standard ordered dither masks of size 130, 130, 128, 144 for the cyan, magenta, yellow, and black
planes respectively. Both on-screen and in-print, the results were sharper halftones. The entire process can be
implemented in parallel and is not computationally expensive.
We investigate design and estimation issues for using the standard color management profile architecture for general custom image enhancement. Color management profiles are a flexible architecture for describing a mapping from an original colorspace to a new colorspace. We investigate use of this same architecture for describing color enhancements that could be defined by a non-technical user using samples of the mapping, just as color management is based on samples of a mapping between an original colorspace and a new colorspace. As an example enhancement, we work with photos of the 24 color patch Macbeth chart under different illuminations, with the goal of defining transformations that would take, for example, a studio D65 image and reproduce it as though it had been taken during a particular sunset. The color management profile architecture includes a look-up-table and interpolation. We concentrate on the estimation of the look-up-table points from minimal number of color enhancement samples (comparing interpolative and extrapolative statistical learning techniques), and evaluate the feasibility of using the color management architecture for custom enhancement definitions.
We consider experimental methods for creating regular grids for applications such as color management where the grids must be estimated from non-grid samples. To estimate the regular grid, we
propose applying a generalization of linear interpolation, called
linear interpolation with maximum entropy (LIME). Evaluating different estimation methods for this problem is difficult and
does not correspond to the standard statistical learning paradigm
of using iid training and test sets in order to compare algorithms. In this paper we consider the experimental issues and propose considering the end goal of the regular grid in evaluating an estimated grid's value. Preliminary experimental results compare LIME, traditional linear interpolation, linear regression and ridge regression.
Wavelet representations of images are increasingly important as more image processing functions are shown to be advantageously executed in the wavelet domain. Images may be inverse halftoned, compressed, denoised, and enhanced in the wavelet domain. In conjunction with other wavelet processing, it would be efficient to halftone directly from the wavelet domain. In this paper we demonstrate how to perform error diffusion in the wavelet domain. The wavelet coefficients are modified by a normalization factor and re-arranged. Then, traditional feed-forward raster scan error diffusion is performed and quality halftones are shown to result. Error diffusing in the wavelet domain is noted to be non-causal with respect to the pixels, and thus the method is not reproducible by feed-forward raster scan error diffusion of pixels. It is shown that the wavelet halftones preserve the average value of the input for constant patches. The resulting halftones may appear smoother in smooth regions and sharper at edges than the corresponding pixel-domain halftones. Disadvantages may include a greater susceptibility to moire and false contouring. Error diffusion is a two-dimensional sigma-delta modulation, and the ideas presented may also be useful for one-dimensional sigma-delta modulation applications.
Image-adaptive color palettization chooses a decreased number of colors to represent an image. Palettization is one way to decrease storage and memory requirements for low-end displays. Palettization is generally approached as a clustering problem, where one attempts to find the k palette colors that minimize the average distortion for all the colors in an image. This would be the optimal approach if the image was to be displayed with each pixel quantized to the closest palette color. However, to improve the image quality the palettization may be followed by error diffusion. In this work, we propose a two-stage palettization where the first stage finds some m << k clusters, and the second stage chooses palette points that cover the spread of each of the M clusters. After error diffusion, this method leads to better image quality at less computational cost and with faster display speed than full k-means palettization.
Single-sensor digital cameras spatially sample the incoming image using a color filter array (CFA). Consequently, each pixel only contains a single color value. In order to reconstruct the original full-color image, a demosaicing step must be performed which interpolates the missing colors at each pixel. Goals in CFA demosaicing include color fidelity, spatial resolution, no false colors, no jagged edges, and computational practicality. Most demosaicing algorithms do well for color fidelity, but there is often a trade-off between a sharp image and the so-called 'zipper effect' or jagged edge look. We propose a novel demosaicing algorithm called vector demosaicing that interpolates missing colors jointly by selecting the color vector that minimizes the sum of distances to the surrounding pixels. The selected color vector is a vector median of the surrounding pixels. The vector median forms an 'average', but preserves sharp edges. We will discuss the theory behind our approach and show experimentally how the theoretical advantages manifest themselves to improve edge resolution while retaining smoothness. Computational complexity is shown to be possible quite low, and we discuss how different approximations may affect the output.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.