A quality metric for HDR (High Dynamic Range)

Gary Demos. CEO of Image Essence, which which is developing wide-dynamic-range codec technology based upon a combination of wavelets, optimal filters, and flowfields, presented on "A Quality Metric for High Dynamic Range. He looked at the Peak Signal-to-Noise Ratio (PSNR) metric, which has long been used for codec evaluation and development, among other uses.

"But the PSNR metric is not suitable for High Dynamic Range," he said. Based on his research, Demos found that  a more appropriate characterization of coding and image quality is to split image brightness into ranges (such as factors of two), and then determine the standard deviation within each such range. Once the standard deviation (sigma) has been determined, the two and three sigma population of pixel differences is shown as percentages of pixels. "This is necessary because codec pixel differences do not typically follow a normal Gaussian error distribution," he said. "The value of sigma at each brightness range, together with the percentage proportions of two and three sigma outliers, provides an appropriate quality metric system for HDR."

In summary, sigma_compare provides a tool for HDR lossy codec characterizations. Outliers are characterized as well as every stop over the range. A self-relative sigma and sigma multiples can be directly compared with the mid-gray perceptual threshold for high quality lossy coding. Knowing self-relative sigma and sigma multiples corresponds approximately to mantisssa bits. But Demos had one caveat. "No automatic metric eliminates the need for visual scrutiny of codec loss patterns."