Tone Mapping and Enhancement
of High Dynamic Range Images

M. Raffin and G. Guarnieri

Site menu:

• Introduction

• Tone mapped images

• 1a. Memorial:
  tone mapping
• 1b. Memorial:
  contrast gain
• 2a. Cathedral:
  tone mapping
• 2b. Cathedral:
  contrast gain
• 3a. Mt. Tam West:
  tone mapping
• 3b. Mt. Tam West:
  contrast gain
• 4a. Napa Valley:
  tone mapping
• 4b. Napa Valley:
  contrast gain
• Results
• References

Introduction

We have developed a new method to display dynamic range (HDR) images on conventional display devices with a realistic rendering of the represented scene. A visual model is applied without the need to specify further information on the represented scene, such as the involved absolute luminance values. The resulting method is computationally efficient and can be easily adjusted by means of only two user parameters.
We also introduce a novel technique to extract the visual adaptation level preserving the sharp transitions in correspondence of the edges. Our approach has been compared with others state-of-the-art tone reproduction algorithms by means of an objective measure which evaluates the gain or loss of local contrast due to the dynamics range reduction.

Overview of the method

Two reasons induced us to develop a new tone mapping operator: firstly, the need of a fast and efficient algorithm, essential for hardware realizations. Secondly, the availability of a new perceptual model which allows to approach the tone mapping problem without the need of detailed information regarding the original scene and the display device.
The problem of tone mapping, in fact, concerns the generation, on a low dynamic range device, of an image which produces a visual stimulus as close as possible to the stimulus produced by the real scene. Hence, an ideal tone mapping operator should consider all the perceptual mechanisms involved during the observation of the real scene and the display. Even if all these visual mechanisms were exactly known, several information related to the environment, the observer state and the real scene would be still required. According to CIECAM model, for instance, in order to accurately specify the conditions in which our eyes perceive an image from a display, we should specify the target image, the observer adaptation state, together with a description of the area immediately adjacent to the display (background) and the remainder of the visual field (surround). Unfortunately, these data are often not available, the parameters might change over the time and their measure would require accurate and expensive instruments. Moreover, the pixel values of the input HDR image are typically expressed in normalized units, and the corresponding physical luminance in the real world are only known up to a scaling factor. In most cases, even if the display minimum and maximum luminance values can be easily measured, it is impossible to know how this interval is positioned compared to the real-world image. Due to these practical limitations, some simplifications are necessary.
In this work we assume that the display and the real world are characterized only by their dynamic ranges. We also assume that the display dynamic range is known, as almost all the display manufacturers provide this element with the device specifications.
The tone mapping operator we propose exploits a global mapping (TRC) in conjunction with a local operator. The TRC allows to preserve the global balance between bright and dim areas while the TRO maintains the perceived local contrast.
The real-world scene, initially characterized by its luminance values, is described in terms of local adaptation level and visible contrast. The local adaptation level is computed efficiently using a novel edge-preserving lowpass filter.
A local dynamic range expansion is introduced to improve the visibility of textures and details with a scheme which resembles the adaptive histogram equalization technique with two substantial differences. In the first place, not all the histogram bins are considered in the computation of the cumulative sum but only those which lie in the interval involved in the perception of the local contrast. Secondly, the cumulative sum does not act directly as an input/output transfer function but operates indirectly influencing the output contrast.

Matlab implementation

The Matlab implementation of our algorithm can be downloaded here.