Metrics for Local-Dimming Artifacts in High-Dynamic-Range LCDs

Local-dimming LCDs exhibit qualities and artifacts that cannot be captured by common performance metrics. For example, a local-dimming display can obtain perfect black levels when the backlight is turned off completely, and the effective measurement of "contrast" will therefore return an infinite value. In this article, robust and meaningful metrics are introduced for the static- and motion-halo artifacts, and good agreement with psychophysical experiments is shown.

by Anders Ballestad, Thomas Wan, Hiroe Li, and Helge Seetzen

THE continued development of light-emitting-diode (LED) backlit liquid-crystal displays (LCDs) has led to the emergence of local-dimming displays, which are entering the marketplace with a promise to deliver high contrast, lower power consumption, and improved image quality. Advanced local-dimming systems also offer increased luminance capabilities for true high-dynamic-range (HDR) imagery. With this broader shift toward dynamic backlighting, it is necessary to consider appropriate performance metrics for such devices.

In principle, any display can be characterized by its external performance characteristics such as peak luminance, contrast, color gamut, spatial resolution, etc. While these specifications can certainly be measured for local-dimming displays, the unique architecture of such displays renders several of them irrelevant. For example, conventional measures of frame-sequential contrast have little value for local-dimming displays because a full-screen black image will result in no light emission by the backlight and thus infinite "contrast." Just as global dimming can return an infinite frame-sequential contrast ratio, local dimming can have a rather dramatic effect on the local contrast, as measured by an ANSI checkerboard. However, due to light scattering in the optical cavity between the backlight modulator and the light-blocking modulator (the LCD panel), the local contrast will most certainly not be as high as the measure of global contrast, and a solid definition of the two should be established and understood.

The unique architecture of local-dimming displays also introduces new artifacts as a result of the dynamic backlight modulation. Conventional metrics do not capture these potential artifacts, and, yet, their impact on the image quality of the display can be considerable if bad design choices are made. In the absence of metrics for these artifacts, any local-dimming configuration achieves very high contrast as a result of the metric limitation outlined above, and comparing different local-dimming designs becomes impossible except by visual inspection. In this article, the most relevant of these novel artifacts are described and metrics for physical characterization of the issues are provided. The results of initial user studies are used to determine the perceived severity of these artifacts for different local-dimming designs. The result of the user studies is important because it provides a perceptual scale over the physically based metric. The result is used to validate the sensitivity of the metric to the measured and perceived artifacts. The weighted metric can then be used to evaluate the perceptual performance of a given display and be used to assist in the design of a new display as a performance design factor.

Local-Dimming Architecture

Before introducing the specific artifacts of local-dimming displays, it is important to understand the architectural differences between local and static backlight designs. For a conventionally backlit LCD, the backlight is of uniform intensity across the entire display area and typically does not vary in average intensity between scenes (global dimming). Therefore, the light-extinction capability of the LCD panel is primarily responsible for making the lowest attainable black level. Local-dimming displays employ an array of addressable light-emitting elements behind the LCD panel. Each light source can be adjusted in intensity across the entire range of full output to no output. Figure 1 shows a sample configuration of a local-dimming display.

The physical layout of the light-emitting elements can vary, and recent commercially available products have ranged from less than 100 to more than 2000 elements for Dolby Vision reference displays (Fig. 2).¹ Likewise, while the light sources of choice are usually LEDs for their appealing environmental and control characteristics, the configuration of the light-emitting element can vary significantly. Local-dimming displays with a large number of elements tend to use a single LED per element, while those with a lower number of elements often combine multiple LEDs into a single block of emitting area. Integration of multiple LEDs in such a design can be achieved through individual wave plates per element or simply by allowing for sufficient diffusion within the optical cavity of the display.

The specific design of the light-emitting element is remarkably irrelevant for the image quality of the local-dimming display. Of course, the choice of design impacts other aspects of the display such as energy efficiency and physical depth of the device, but only the spatial and angular distribution of light emitted by the element is relevant for image quality. The local-dimming array can therefore be described by the pitch between individual light-emitting elements and the point-spread function (PSF) of light emitted by the element. The concept of a PSF still applies even if the light-emitting element is a much larger structure composed of multiple LEDs because the easiest representation for such an arrangement is still just the positioning of individual (though possibly complex) PSFs in intervals given by the center-to-center pitch of the array.

Different algorithmic solutions can be used to drive the light-emitting arrays but, in general, the drive values for the elements are obtained from the corresponding local image data. The LCD image is then adjusted in some fashion to compensate for the variable low-resolution light field generated by the light-emitting elements under those drive conditions. The form of compensation can vary from a very general estimate to a detailed mathematical prediction of the light field. Independent of the specific choice of algorithm, it is important to understand that the compensation by the LCD for the low-resolution variation on the light-emitting array is a critical part of the local-dimming design. No or incomplete compensation will exaggerate the local-dimming specific artifacts significantly as described in the following sections.

Fig. 1: Typical local-dimming architecture. The point-spread function is the light profile provided by each LED (or light element) onto the LCD.

Fig. 2: SIM2 Dolby Vision local-dimming display.¹

Algorithmic compensation as described above can only succeed if enough light is generated by the backlight in each region of the image. This is particularly challenging for large regions of high brightness where the backlight needs to be uniformly bright. To achieve this condition, the PSFs from neighboring light elements need to overlap spatially so that no lower luminance gaps appear between light elements. This solution addresses large, bright areas but can lead to complications for small, bright features. When displaying small, bright objects on black backgrounds, the generated backlight can be larger than the intended pattern itself, and the finite-contrast panel cannot hide the excess light, resulting in the appearance of a cloud or halo of light around the object.

An example of particularly difficult content for the PC application of local-dimming dis-plays is the ubiquitous mouse pointer. Scrolling movie credits and Microsoft Windows^® "star-field" screen saver are other common examples of content that would suffer equally from this artifact. When viewed on a black background, the LCD cannot compensate for the light that leaks through the finite-contrast panel around the intended bright pattern; for all other non-zero background gray levels, the artifact can be removed by compensating for it on the LCD.

Static Halo

The halo is in effect an unwanted cloud of light around a given intended pattern. The halo is only noticeable if it is of intermediate extent, i.e., it cannot be observed if it is really small, or really big, but this latter case is in effect a flat-backlight reduced-contrast display and not a local-dimming display. The severity of the halo can for small PSFs be described by the following expression:

halo metric = total halo luminance / total image luminance.    (1)

This expression will not capture the extreme part of the spectrum where the backlight is flat, so a correction term will be necessary if one is to estimate the halo for very large PSFs; for example, (1-A^m )ⁿ, where A is the ratio of light in the measured halo relative to an infinitely extended halo (which then equals a flat-backlight display) and m and n are fitting parameters. Typically, however, the filling factor A is quite small for any reasonably sized PSF.

The shape of the halo is also important, as is its center of mass relative to the intended pattern, but the severity of the halo artifact is typically captured consistently by the expression given by Eq. (1). In order to perform this calculation, an image of the entire display was collected using a luminance-imaging camera, such as a Lumetrix 400A imaging photometer system.² The halo test image displayed was a small circle, which is representative of a small feature of interest that may cause a halo. Because the metric aims to quantify the sever-ity of the artifact for any given display and any given backlight-generating algorithm, these need not be specified. A luminance image taken under the prescribed experimental conditions on a local-dimming 37-in. 1080p display with 1380 individual LED light elements spaced approximately 19 mm apart is shown in Fig. 3(a), with a horizontal cross-section shown in Fig. 3(b) (green line), both plotted on a log scale.

(a) 

(b) 

The luminance image presented to the camera is adjusted to construct the true physical display halo. To achieve this, three luminance images are captured of the same test pattern at maximum, low, and lowest LCD panel trans-mission with the same backlight intensity level for all three. Using the difference between the three captures, the effect of camera scatter can be isolated and the true display halo and also the boundary between the halo and the test pattern can be determined. Both the true halo (blue line) and the test pattern (green line) are shown in Fig. 3(b), and the scattering in the test pattern is obvious.

Motion Halo

While having a halo is not necessarily desirable, much of it can be covered up by the fact that observers are used to it in the form of veiling luminance, or scattering in the eye.⁴ This puts an upper limit on just how many backlight elements are necessary on the backlight which can be readily calculated using known veiling-luminance models.^5,6 For example, if a halo with a full-width half maximum (FWHM) of 1 in. is just hidden by veiling luminance, then one would need on the order of one light element per square inch of display area.

However, if the halo is visible, i.e., if it is larger than what can be hidden by veiling luminance, then a temporal change in its size or shape can be noticeable. Furthermore, if the test pattern moves small distances, but the halo stays put (as it is related to the static position of the light elements), then the relative center-of-mass difference of the pattern and the halo will also change. This will result in the illusion of the halo "walking" or "wobbling" along with the smoothly moving test pattern.

The following describes a test metric for this motion-halo artifact. A 22-pixel-radius white dot is set against a black background. The photometer is placed in a stationary position perpendicularly 2 m away from the center of the display and images the dot as it traverses a 300-pixel-radius path around the center of the display. (The photometer settings were set to f/5.6 with a focal length of 12.5 mm, which makes the photometer aperture about 2.2 mm across.) Figure 4 shows the result of these measurements on the same 37-in. locally dimmed display that we discussed above. The mean halo metric was found to be 0.0050 and its standard deviation 0.0007. The FWHM/2 of the PSF of this particular display is 33 LCD pixels.

The motion- halo metric is described as being the ratio of the halo-metric mean to the standard deviation. Therefore, the motion-halo artifact can be detected by simply calculating the static-halo metric for a series of successive still images. For the example shown, the percentage variation in the motion-halo metric is 0.0007/ 0.0050 x 100 = 14%.

Both of these metrics provide a numerical technique to measure halo artifacts. They are easy to execute with conventional test devices and relatively insensitive to small measurement error. The final step is to evaluate the perceived quality impact of these artifacts.

User Studies on Static- and Motion- Halo Artifacts

Because the number of local-dimming displays in the marketplace is still small, a flexible simulator is used to map out the range of light-element configurations expected in the marketplace. This simulator system comprised a high-luminance projector whose image was relayed onto the back of a conventional 40-in. 1080p 1000:1-contrast-ratio LCD panel, both having a refresh frequency of 60 Hz. If the images on these two spatial modulators are synchronized, then the projector image can be used to simulate the light field of a local-dimming display.³

The severity of the halo artifacts was studied by using two methods. Ratio-scaling was used to map out the general user response to the static-halo artifact for halo sizes ranging from non-existent to flat backlight. In a second set of experiments, the method of constant stimuli was used to find the threshold for both the static- and the motion-halo artifacts. Participants sat in a dark room at a distance of 3 m in front of the display system when the experiment was performed. Fourteen participants completed the user studies (the average age was 31 years old, and there were nine female participants). The dot size was changed to a radius of 10 pixels in order to minimize the effects of veiling luminance, and the following results are therefore not directly comparable with the experiment in the previous section. The general shape of the simulated backlight was obtained by fitting it to a measurement of a PSF from an HDR display with 1380 light elements. The lateral extent of this PSF was then varied. The results from the static-halo user response are shown in Fig. 5. The threshold for the static-halo artifact under these experimental conditions was found to be at a PSF size of 20 ± 9 pixels (FWHM/2), for which the corresponding A parameter is indicated by the black vertical line in Fig. 5.

Fig. 4: Results from the motion-halo-artifact measurements on a 37-in. local-dimming display with 1380 LEDs. The graph shows the halo size for 300 successive images taken of a dot with a radius of 22 pixels traversing a larger radius (a 300-pixel circle) in a clockwise orientation around the center of the display. The mean halo metric is 0.0050 and its standard deviation is 0.0007.

We have also plotted the expression in Eq. (1) in Fig. 5, both with (solid line) and without (dashed-dotted line) the large halo correction term (1-A)², where the constant "2" was obtained by fitting. Veiling luminance is also included in this fit by adding a constant 0.6% of the contribution from a flat backlight to both the numerator and denominator of Eq. (1). For all reasonable halo sizes, the correction term for large halos is not necessary, and the expression in Eq. (1) alone is adequate.

For the user studies investigating the motion-halo experiments, the measurements displayed in Fig. 4 were mimicked by moving a small dot around in a big circle, and allowing for the halo size to vary along the way. A mean halo size of 45 pixels (FWHM/2) was used. The motion halo was found to be more visible the larger the amplitude of the oscillation got, and a threshold of 2.7 ± 0.8 pixels was determined. This means that the user will observe the artifact if the motion-halo metric exceeds 6% (oscillation amplitude divided by mean halo size: 2.7/45). In this experiment, the halo size oscillated at 2 Hz. For larger frequencies, i.e., faster-moving features, there will be a cutoff where the artifact is no longer visible due to under-sampling.

Conclusion

A method for characterizing the static- and motion-halo artifacts in locally dimmed displays has been outlined and metrics developed. Psychophysical experiments verified the expected functional form and, furthermore, produced thresholds for both artifacts under a given set of test conditions. The static and motion-halo artifacts are fundamentally related to the architecture of local-dimming displays. Metrics sensitive to the specific artifacts enable the designer to optimize for the desired quality and avoid poor performance, as even the smallest halo can result in a wobble effect, or conversely, a larger halo can remain unseen if its size remains largely constant.

References

¹Dolby Laboratories, Inc.; www.dolby.com.

²Lumetrix Corp.; www.lumetrix.com.

³H. Seetzen, L. A. Whitehead, and G. Ward, "A High Dynamic Range Display Using Low and High Resolution Modulators," SID Symposium Digest 34 (2003).

⁴M. Trentacoste, "Photometric Image Processing for High Dynamic Range Displays," M.Sc. thesis, University of British Columbia, Canada (2006).

⁵G. Spencer, P. Shirley, K. Zimmerman, and D. P. Greenberg, "Physically Based Glare Effects for Digital Images," Computer Graphics 29, 325-334 (1995).

⁶CIE (International Commission on Illumination), "CIE Collection on Glare, 2002." •

Fig. 5: Results from ratio-scaling user studies for the static-halo artifact. The dashed-dotted line shows the expression from Eq. (1), and the solid line shows Eq. (1) multiplied by the large halo correction term (1-A)². The threshold for the static-halo artifact was determined to be 20 ± 9 LCD pixels (FWHM/2). Dolby Vision is a high luminance (1500 cd/m² or higher) high-LED-density display design, and Dolby Contrast is a more conventional (up to 650 cd/m²) medium-LED-density design.