Incremental Color Quantization for Color-Vision-Deficient Observers Using Mobile Gaming Data

I account for a large fraction of the data transmitted on the Internet (1). The sizes of these images, when compressed for storage and transmission, depend on their spatial resolution (number of pixels) and color resolution (number of color quantization levels). Existing image compression methods such as JPEG take into account human visual perception but assume an individual with normal color vision (2). Because compression algorithms can provide improved compression when they target a smaller number of distinguishable image colors, these algorithms could in principle use information about color vision deficiencies in observers to improve compression ratios. Approximately 8% of males and about 0.5% of females have color vision deficiencies (3–5). There is therefore a missed opportunity to harness information about color vision deficiencies to improve image compression and thereby to deliver improved network speeds and improved file storage efficiency. Users of mobile devices spend over 70% of their time in applications such as web browsing and productivity, as opposed to gaming and multimedia (6). This observation, together with the personal usage model of mobile devices and the popularity of user-specific sites and web applications therefore makes it practical to transcode images on a server to deliver improved performance customized to individual users. Today, the fraction of almost 9% of mobile device owners with color vision deficiencies is missing out on the opportunity for improved network performance and reduced data storage requirements. A. DaltonQuant: Higher compression rates by exploiting human color vision deficiencies. We introduce DaltonQuant, a new method for bespoke (user-specific) image compression. DaltonQuant uses incremental color quantization to improve image compression for observers with color vision deficiencies, reducing file sizes by up to 29% more than the state of the art. DaltonQuant builds a user-specific model for the color acuity of its target users. This model characterizes the colors which each target observer considers to be visually equivalent. We provide two different instantiations of DaltonQuant’s userspecific function for quantifying color equivalences, constructed from a dataset of 28 million color comparisons performed by humans on modern mobile displays. The dataset used was collected through the mobile game Specimen (7). We use an analysis of the data collected by the Specimen game, across all its users, to identify individuals in the mobile game data who demonstrated game behavior consistent with color vision deficiencies. We use the data from 30 of these individuals to construct and evaluate the two concrete instantiations of our technique. We evaluate compression performance for these 30 individuals and show that DaltonQuant enables average improvements of 22%– 29% file size reduction (i.e., a compression ratio of 1.28 – 1.40) over state-ofthe-art compressors TinyPNG and pngquant for the Kodak PC image benchmark suite. DaltonQuant’s improvements result from undoing conservative decisions made by TinyPNG and pngquant, neither of which customize compression for observers with color vision deficiencies. Color quantization reduces the number of colors used to represent an image and is a standard component of mainstream lossy image compression techniques (8–10). To avoid unacceptable image degradation, successful color quantization techniques replace a large set of existing colors with a smaller set that is perceived by most human observers to be similar to the original set. All existing color quantization techniques assume a viewer with normal color vision and no vision deficiencies when constructing their reduced color palette. Because an image compression algorithm may fail to combine colors that an observer would not easily distinguish, this assumption leads to a missed opportunity to provide improved compression ratios for color vision deficient observers.