Purchase this article with an account.
Yuanzhen Li, Lavanya Sharan, Edward H. Adelson; Perceptually based range compression for high dynamic range images. Journal of Vision 2005;5(8):598. doi: 10.1167/5.8.598.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Natural scenes contain huge ranges of luminance, and it has recently become convenient to capture high dynamic range (HDR) images digitally. Unfortunately there is no easy way to display them. The classical way of handling dynamic range in photographs is through a point non-linearity, such as a log or power function. However, if we forbid clipping of highlights with an HDR image, this forces a loss of contrast, especially in shadows. Local gain control techniques can help, but they lead to an unnatural appearance where edges become “cuspy.” More sophisticated techniques have recently been proposed, but they have disadvantages in terms of computational speed or visual quality. We have developed an alternate technique inspired by human vision. An image is split into subbands, and local gain control (i.e., contrast normalization) is applied to each subband, and then the image is resynthesized. This works remarkably well, allowing huge compression while keeping detail in both bright and dark regions and avoiding unnatural artifacts. Why should it work? We believe it solves the following problem statement, which is parallels that used in color: Given the limits of a display device, choose the displayable image that causes the relevant neural responses to most closely match the neural responses to the original scene. We say “relevant neural responses” because we can't make the retinal responses match those for the original scene, but we can try to attain approximate matches for later stages such as area V1. We can do this by applying gain controls in the luminance and subbband domains, chosen so that they will be cancelled out by the gain controls in the retina and cortex. This gives us the degrees of freedom we need to best satisfy our conflicting constraints. In the present case, it means that we can greatly compress the dynamic range of an image, allowing it to be displayed on a monitor or printed on paper, while retaining detail and avoiding irritating artifacts.
This PDF is available to Subscribers Only