Current display technology has a very limited dynamic range, so if an HDR image is to be displayed, the dynamic range of the image will have to be compressed to that of the display device. This process is known as tone-mapping.

Tone-mapping operators are divided into four categories. More details on tone-mapping and all aspects HDR related may be found in Reinhard et al: High Dynamic Range Imaging - Acquisition, Display and Image-Based Lighting. The link is listed on the References page.

Global operators
The same (non-linear) curve is applied to all pixels,
Local operators
The adaptation level is derived for each pixel individually considering the local neighbourhood,
Frequency domain operators
The dynamic range is reduced based on the spatial frequency of the image region,
Gradient domain operators
A derivative of the image is modified.

A number of operators exists that try to mimic the human visual system. The human vision is a fairly complex process with a highly non-linear response. Those tone-mapping operators require the image to be in real-world units, i.e. cd/m². Because the colour-sensitive cones work for high illumination, with the rods being responsible in low-light environments, one of the effects of human tone-mapping operators is the desaturation of colours for low luminances. Similarly, very bright regions of the image will result in a veiling luminance thus causing glare.

LDR auto-exposure
Auto-exposed LDR photograph for comparison

Many different tone-mapping operators exists, and the optimum operator for a particular application and output device might require some experimentation from the user.

Global Operators

The simplest possible tone-mapping is linear mapping, but because most display devices exhibit a non-linear response, this will result in very dark images if the HDR image has a wide dynamic range. A gamma-corrected linear mapping attempts to correct this problem. Exponential and logarithmic corrections to linear mapping are also possible.

Drago extends extends the logarithmic response curves in order to handle a wider dynamic range. A logarithmic compression is applied to the image luminance. The base of the logarithm is varied between 2 and 10, based on the brightness of regions within the image. This results in a preservation of contrast in darker regions and a higher compression for bright regions.

Drago 03 Reinhard 02
Drago 2003 (left), Reinhard 2002 (right)

Reinhard et. al. took the inspiration to their tone-mapping operator from techniques known from traditional wet-film photography. To get the visually best print from a negative, it is not enough to just match contrast and brightness. Scene content, image medium, and viewing conditions must often be considered, too. It was Ansel Adams who developed the so-called Zone System in the 1940s. Now, 60 years later, it is still used successfully because it combines quantitative measurements with artistic image content.

Local Operators

The Pattanaik multi-scale observer operator attempts to model all steps within the human visual system currently known well enough to be modelled. It is the most complete framework to date, although not strictly speaking necessary for only reducing the dynamic range.

Ashikhmin 02
Ashikhmin 2002

In contrast to the Pattanaik model, Ashikhmin's operator only implements those aspects relevant to dynamic range compression. As a result, the Ashikhmin operator is significantly faster.

Frequency Domain Operators

An image may be thought of as being composed from a LDR component with a high spatial frequency, and a HDR component for low frequencies. If the two can be separated, only one of the components has to be compressed. This concept has been implemented by Durand and Dorsey. It is based on bilateral filtering which is an edge-preserving operation that blurs the image while keeping edges intact.

Durand 02
Durand 2002

Gradient Domain Operators

While high-frequency components within an image cause rapid changes in luminance and high gradients, those are much smoother with low-frequency objects. The gradient field, i.e. a representation of the pixel changes, is compressed with operators such as Fattal's gradient domain compression. By varying the amount of compression, the amount of details that is preserved in the displayed image may be adjusted.

Fattal 02
Fattal 2002

For the research papers, see References