Color-to-gray tested techniques

For our context we analyzed several algorithms to cover a wide range of approaches. Concerning Image Space conversions, we chose the Green-to-Grey as trivial method and the MATLAB RGB2Gray as direct method because of its relationship with hu- man vision and its popularity in the Computer Vision community.

Between direct methods we tested also Adobe Photoshop conversion using prede- fined settings. Adobe Photoshop devised also custom non-linear projections, but these require users to set image-dependent parameters by trial and error [12]. For this reason were left out of our tests.

Between Chrominance direct methods, we discarded all methods tested accurately without success in [6] and also the methods discarded by this study with consistent motivations. E.g. Gooch Color2Gray [13] not was implemented because other expe- riences demonstrate that, although its gradient-preserving nature could improve fea- tures discriminability, in practice it does not improve the quality of the results be-

cause of its inherent problems with the input parameter selection and its inconsistent spatial locality. We implemented the local methods Smith et al. [9] technique but we abandoned them soon due the problems in our MVS software (nFrames SURE) [14] where its adaptively-weighted multi-scale unsharp mask generate large problems (no models produced in the ground dataset): it’s well known that the unsharp masking filter enhances (modify) the fine details of the image and colors are mapped incon- sistently between different parts of the images depending on the surrounding neigh- borhoods. While the algorithm can use spatial information to determine the map- ping, the same color should be mapped to the same grayscale value for every pixel in the image. Between Chrominance direct methods we implemented Grundland and Dodgson [11] and two techniques of the same authors: Contrast Preserving Decolor- ization [15] and Real-time Contrast Preserving Decolorization [16]. Both based on the same theoretical framework and algorithm, they differ only for simplification technique introduced with the latter.

We found that in our case the second solution is more suitable for two reasons: a. a much shorter and constant (independent from resolution) processing time

(less than one second);

b. Contrast Preserving Decolorization, to be processed in real time, implements a

simplified version of the original equation. This version has the beneficial drawback to convert the patches appearing in different images univocally.

Preliminary tests with both the methods prove the efficiency of the second one. A detailed description of the chrominance direct methods tested follows. Grundland and Dodgson

Grundland and Dodgson [11] performed a global grayscale conversion by expressing grayscale as a continuous, image-dependent, piecewise linear mapping of the pri- mary RGB colors and their saturation. Their algorithm, called Decolorize, works in the YPQ color opponent space and aim to contrast enhance. The color differences in this color space are projected onto the two predominant chromatic contrast axes and are then added to the luminance image. Unlike principal component analysis, which optimizes the variability of observations, predominant component analysis optimizes the differences between observations. The predominant chromatic axis aims to cap- ture, with a single chromatic coordinate, the color contrast information that is lost in the luminance channel. The luminance channel Y is obtained with the NTSC CCIR 601 luma weights. Grundland Decolorize is very sensitive to the issue of gamma compression with some risks of decrease of the quality of the results mainly in light areas or dark areas where many features will be lost because the saturation balancing interacts incorrectly with the outlier detection.

Real-time Contrast Preserving Decolorization

The human visual system does not univocally perceive chrominance and lightness, while their relationship to the adjacent context plays a vital role and that the order of different colors also cannot be defined uniquely by people, Lu et al. [13] relax the color order constraint and present a new method seeking better preservation of color contrast and significant enhancement of visual distinctiveness for edges. For color pairs without a clear order in brightness, authors propose a bimodal distribution, i.e., mixture of two Gaussians, to automatically find suitable orders with respect to the

visual context in optimization. This strategy enables automatically finding suitable gray scales and preserves significant color change. Practically they use a global map- ping scheme where all color pixels in the input are converted to grayscale using the same mapping function (a finite multivariate polynomial function). Therefore, two pixels with the same color will have the same gray scale. The technique is today im- plemented in OpenCV 3.0. In order to achieve real-time performance, authors fur- ther devise a discrete searching optimization which takes advantage of a linear par- ametric grayscale model as well as a sampling based P-shrinking process [16]. Spe- cifically, they approximate their previous optimization-based method and achieve real-time performance by confining the polynomial color model into a constrained, discrete linear color model. To further speedup the decolorization process, they down-sample the high-resolution input to a small scale 64 × 64. This is valid due to the inherent color redundancy of natural images. Extensive experiments show that the proposed P-shrinking scheme can achieve real-time performance for high- resolution images, without obvious quality degradation. Also if the approximated so- lution in confined search space might produce unsatisfactory results in special cases this last solution is suitable for our case solving the problem using fixed strategies not linked with perception but with a simple clear order in brightness, also if the above-mentioned methods do not consider the preservation of brightness of the input color image. Main drawback is that in two different images same color could be converted in different grays.