Evaluation of floating point image compression

Recently, compression of high dynamic range (HDR) photography gained attention in the standardization of the Microsoft HDPhoto compression scheme as JPEG-XR. While integer data of 16 bits/pixel (bpp) in scRGB color-space can represent images up to a dynamic range of about 3.5 magnitudes in luminance - a noteworthy improvement over the 1.6 magnitudes possible in sRGB - even higher ranges are more efficiently represented by floating-point number formats. However, traditional means to evaluate image quality are rarely suitable for such data: They are often only calibrated to low dynamic ranges (LDR) of 8 bpp, and are not designed to take the peculiarities of floating-point data into account. In this work, we present two approaches to deal with this problem by introducing a (mathematical) quality index more suitable to floating point data related to SSIM and by presenting an independent image quality evaluation framework that is able to apply LDR metrics to HDR data. The presented ideas are then tested on the HDPhoto floating point compression, on a proprietary backwards compatible extension of JPEG and on a proposed floating point compression scheme based on JPEG 2000 that is also proven to be optimal in the proposed quality index. It is then shown that both approaches, the proposed metric and PSNR in the LDR domain, deliver comparable results.