Engineers from Intel have introduced the Spectral JPEG XL image format, optimized for efficient compression of images covering spectral regions beyond the visible spectrum. The proposed format offers capabilities similar to the spectral edition of the OpenEXR format, but unlike the latter, it provides lossy coding, which allows for a reduction in file size by 10-60 times compared to lossless compression.
Spectral images contain not only light intensity information in the three primary color channels (RGB), but also cover parts of the ultraviolet and infrared ranges. Such images are used in areas such as high-end rendering, material analysis, and scientific data visualization. For example, spectral images can be used to accurately model real-world optical effects in rendering, to assess how paint appears under different lighting, and to identify materials based on their light signatures.
Shooting beyond the visible spectrum allows for more accurate modeling of how light interacts with materials, but at the cost of significantly more information stored for each pixel. Spectral images can include dozens of channels covering different wavelength ranges, and use 16- or 32-bit floating-point numbers for each channel, allowing for a wider range of brightness values than in conventional photographs. The price of this capability is a significant increase in size compared to traditional images. An example is an image with 81 additional spectral channels that takes up 300 MB in OpenEXR format. Using Spectral JPEG XL, this image was compressed to 3.9 MB without losing spectral characteristics.
To reduce the size, Spectral JPEG XL uses separation of brightness and spectral shape data, and applies a discrete cosine transform, which allows preserving the main spectral characteristics but discarding unimportant high-frequency spectral details. The essence of the method is to transform a smooth change in wave characteristics into a set of wave coefficients (frequency coefficients), which, when combined, recreate the original spectral information.
The higher frequency spectral coefficients are then normalized to the overall luminance, allowing less important data to be more aggressively compressed. The idea is that the average luminance is the most perceptually important and is preserved at the highest quality, while the high frequency coefficients are less important and are subject to higher compression levels and optionally lower resolutions. The information is then processed by a codec built on the existing compression engine developed for the JPEG XL format.
Source: opennet.ru
