Which tone-mapping operator is the best? A comparative study of perceptual quality

Abstract

Tone-mapping operators (TMOs) are designed to generate perceptually similar low-dynamic-range images from high-dynamic-range ones. We studied the performance of 15 TMOs in two psychophysical experiments where observers compared the digitally generated tone-mapped images to their corresponding physical scenes. All experiments were performed in a controlled environment, and the setups were designed to emphasize different image properties: in the first experiment we evaluated the local relationships among intensity levels, and in the second one we evaluated global visual appearance among physical scenes and tone-mapped images, which were presented side by side. We ranked the TMOs according to how well they reproduced the results obtained in the physical scene. Our results show that ranking position clearly depends on the adopted evaluation criteria, which implies that, in general, these tone-mapping algorithms consider either local or global image attributes but rarely both. Regarding the question of which TMO is the best, KimKautz ["Consistent tone reproduction," in Proceedings of Computer Graphics and Imaging (2008)] and Krawczyk ["Lightness perception in tone reproduction for high dynamic range images," in Proceedings of Eurographics (2005), p. 3] obtained the better results across the different experiments. We conclude that more thorough and standardized evaluation criteria are needed to study all the characteristics of TMOs, as there is ample room for improvement in future developments.

Perceptually optimized image rendering

“…Nowadays, tone mapping methods often do not make explicit use of perceptual metrics (see [5] for a nice review), but rather provide the user with a small set of free parameters to hand-control the mapping from the scene to the displayed image. These methods are conceptually simpler than ours, and some of them can produce high quality results in controlled situations (see for instance [21]).…”

Blind Tone-Mapped Image Quality Assessment Based on Regional Sparse Response and Aesthetics

“…To obtain the perceptual features in the sparse domain, a novel TMI training dataset is constructed as the basic of multi-dictionary learning. Specifically, we selected 20 pristine HDR images with different kinds of scenes from existing HDR image datasets [ 31 , 32 ] and generated the corresponding distorted versions processed with 15 classical TMOs [ 33 ]. To avoid the same TMI datasets used in dictionary learning and quality assessment stages, the image contents contained in TMI training dataset were distinct from those in the subsequent-used benchmark database for objective quality assessment (i.e., ESPL-LIVE HDR database [ 34 ]), and Figure 2 depicts the partial scene contents, which includes the indoor, outdoor and night scenes.…”

Tone Mapping Operators: Progressing Towards Semantic-Awareness

“…Global TMOs apply the same luminance compensation throughout the image, whereas local TMOs take into account the spatial neighborhood of each pixel. We follow a recent comparative subjective study of several classical TMOs provided by Cerda-Company et al [4] to understand the performance of different TMOs. Kim et al [2] (rated highly over subjective experiments [4]) propose a global TMO based on the luminance adaptation of human visual cortex.…”

“…We follow a recent comparative subjective study of several classical TMOs provided by Cerda-Company et al [4] to understand the performance of different TMOs. Kim et al [2] (rated highly over subjective experiments [4]) propose a global TMO based on the luminance adaptation of human visual cortex. They suggest that human visual sensitivity is adapted to the average log luminance of the scene and that it follows a Gaussian distribution.…”

“…As a local approach, Krawczyk et al [1] propose a TMO based on a probabilistic model of lightness perception (rated highly over subjective experiments [4]). They decompose an HDR image into areas of consistent luminance (lightness framework) and map each framework by adjusting the perceived 'white' point based on Gilchrist's anchoring rule [5].…”