This page describes how the LCD settling behavior is specified and measured. Regarding the set of performance measures, we distinguish between two LCD operating modes: Strobed backlight (LightBoost and similar) and flicker-free backlight. This page is only about the settling performance in the flicker-free backlight mode – for the settling performance in the LightBoost mode see LightBoost settling.
- Side note: Both modes, flicker-free and strobed backlight, can be equally useful, but it all depends on the application. Strobed backlight is certainly useful when it comes to presenting motion stimuli and avoiding motion blur, no matter whether the stimuli are possibly tracked with the eyes. Strobed backlight also offers a basically instant and fully synchronous stimulus onset and offset, which comes, of course, with flicker that is usually not meant to be part of the stimulation. Synchronous stimulus onset means that the stimulus appears at all screen locations at the same time (sufficient settling performance assumed).
- Flicker-free backlight, on the other hand, is useful for static stimuli or for stimuli which are animated in place, meaning for stimuli which do not move but are possibly switched on and off at fixed locations. The stimulus onset and offset is not as instant as with strobed backlight, and the screen is not updated all at once but, instead, from top to bottom. However, there is no flicker unless flicker is intended to be part of the stimulus.
- Actually, both backlight modes have aspects that are similar to the operation mode of the good old CRTs. The pulse-like excitation of the phosphors is more similar to the strobed backlight mode, whereas the line-wise screen refresh is more similar to the flicker-free backlight mode.
The goal behind representing the settling behavior by a few graphs and numbers is to quantitatively rate the performance of a monitor and to compare it with other monitors or the requirements of the application at hand. It is difficult, however, to come up with a set of performance measures that is small, easy to understand, easy to measure, and of practical relevance. What is practically relevant depends, of course, on the application at hand and on the properties the monitors actually can differ in. Regarding the latter, an overshoot measure, for example, does only make sense if the monitors actually differ in overshoot behavior, which they only do since overdrive technologies have been implemented. So to some extent, the set of performance measures needs to be adapted to the ever changing monitor technology. Using an inappropriate set of performance measures is not only misleading us, the customers, when looking for a good monitor but might also make manufacturers optimize monitors in the wrong way.
In order to quantify things as objectively as possible, the numbers should be based on physical measurements. However, what matters in the end are not the physical properties directly but their impact on our perception as the signal travels through the visual system. Unfortunately, quantifying perception and inferring perceptual quantities from the physical properties is not as trivial as it might seem. We know a few things about the visual system, but this knowledge is derived from rather specific experiments that were run under well controlled lab conditions and do not necessarily translate well into real world scenarios where many more factors play a role. One example, which is also relevant for monitor characterization, is Weber's law, which basically says that the visual system's sensitivity for luminance differences scales with the luminance level at which the differences are observed. Sensitivity for a luminance change is higher for dark stimuli than it is for bright stimuli. However, in how far Weber's law holds for a given spatial stimulus configuration or even for dynamically changing scenes cannot be easily inferred just from the simple case/experiment for which Weber's law has been found true. Even if we had an accurate model of how the test stimuli are processed by the visual system, we still would not know how well the test stimuli adopted for the measurement taken here are representative for arbitrary image sequences in real world applications.
All analysis was done for a low-pass filtered luminance signal, using a Gaussian low-pass with a -3 dB corner frequency of 70 Hz. 70 Hz is assumed to be close but still safely above the critical flicker fusion frequency (CFF). It is worth mentioning that the CFF is not a hard limit beyond which humans cannot detect flicker anymore. Likewise, the corner frequency of a Gaussian low-pass filter is not a hard limit beyond which no energy can pass the filter anymore. Anyway, applying such a low-pass filter makes the analyzed signal look more like it is "seen" by the visual system. Moreover, low-pass filtering increases the S/N ratio, thereby making analysis easier and more robust. On the other hand, such low-pass filtering might be the limiting factor when it comes to measuring very small rise/fall times. For example, filtering an ideal step function signal with a 70 Hz Gaussian low-pass filter would result in an apparent rise time of about 4.8 ms.
The luminance levels used for the step matrix measurements were uniformly distributed across the pixel value space, and the monitor was calibrated to a gamma value of 2.2. Level 0 corresponds to black, level 8 corresponds to 100% white. Using a gamma different than 1 means using a non-linear luminance spacing. This causes the difference in physical luminances between Level 0 and 1 to be way smaller than the difference between level 7 and 8, whereas the perceived differences between these luminance steps should appear much more similar, according to Stevens' law (or similar laws). See also Luminance scaling & settling errors.
Each bar in a matrix plot represents the average of N=5 measurements of the shown performance parameter when going from luminance level "FROM" to level "TO". Only absolute (i.e., positive) values are presented, but where applicable, the sign of the underlying signed values is indicated by the colors of the bars (red codes for negative). The thin green bars, if not too small to be visible, show the variation of the measure and extend by one standard deviation above the corresponding main bars. Do not confuse the standard deviation with the standard error, which is, for N=5, only about 45% of the standard deviation (45% ≈ 1/sqrt(N)).
The gray and flat bars drawn on the back walls of the coordinate system cube are the projections of the main bars along the FROM and TO direction respectively. For some plots, bars more in the front of the plot might completely occlude bars behind them. Where this is the case, an additional plot might be provided which shows the matrices from a different angle.
For each matrix plot, two maximum values are provided, one being calculated by taking 100% of all step sizes and directions into account (without counting the FROM=TO cases along the diagonal where the step size is zero), and one taking only the best 80% into account, i.e., leaving the worst cases unconsidered.
The settling time is defined as the time after which the luminance curve stays within the ±10% or ±2% error band. This time is given with respect to when the luminance starts changing, an event not directly measured but inferred from special triggering frames (see Triggering section below) so as to avoid that the reference time depends on the particular luminance step size.
The error bands (±10% or ±2%) are defined relative to the luminance step size, which is the way suggested by the Information Display Measurements Standard (IDMS). See also Luminance scaling & settling errors for an in-depth discussion of this topic.
The delay refers to the time the signal needs to reach the 50% point (i.e., half of the step size) and is somewhat related to the slope of the rising or falling edge (see below). However, the delay measurement also captures dynamic effects at the very beginning of the settling curve, for example, additional delays caused by the LC cells being in a saturated state. Moreover, the 50% point can serve as a reference point for the effective onset or offset time of a stimulus. The reported values are deviation values though, given with respect to the average delay taken over all measured step sizes. The delay deviations are always reported as absolute (i.e., positive) values, but the original sign of the delay is color coded in the plots (red coding for negative).
It is desirable to have the variance of the delays across all luminance step sizes minimized, because in colored scenes each color channel might "see" different luminance switching conditions, causing color shifts during the settling phase if the settling behaviors are too different. Obviously, the same holds true for other performance measures, like the settling time for instance.
Presuming a low-to-high luminance change, overshoot is the difference between the maximum of the luminance curve and the settled luminance, reported as percent of the luminance step size. For a high-to-low change, we basically switch the definition and just use the minimum instead of the maximum and invert the sign of the difference so that the overshoot value is always zero or positive, even for high-to-low changes.
There can only be undershoot if there was overshoot, otherwise the undershoot will be zero. If there was overshoot, say for a low-to-high luminance change, the undershoot is the difference between the minimum of the luminance curve and the settled luminance. Only the luminance curve after the overshoot time point is taken into account when looking for undershoot. The reported undershoot values are given in percent of the luminance step size.
The rise or fall time is defined by the time the signal needs to go from 10% to 90% (or vice versa), where 100% is the individual nominal luminance step size. Obviously, short rise and fall times are good, but the times should also be as homogenous as possible across different step sizes and step directions (rise and fall), for reasons explained for the Delay deviation measure already.
Be aware that the measured rise/fall time is potentially limited by the applied 70 Hz low-pass filtering to about 4.8 ms. Values falling far below this limit would indicate rather excessive overshoot in the original signal.
Frame luminance error
The frame luminance error is the deviation of the average luminance from the target luminance (i.e., the luminance in the settled state), where the average is calculated over exactly one refresh period. We let the first refresh period begin at the average delay (see above, Delay deviation), i.e., at the average 50% point of the luminance step curves. With this choice we kind of minimize the average luminance error for the first frame implicitly. The errors are reported as percent of the particular luminance step size.
Note that the errors for the first refresh cycle are not that relevant in the flicker-free mode as they might just reflect the choice of the frame binning in conjunction with the form of the onset/offset luminance curves. More important might be the error distribution, which goes hand in hand with the delay deviation. Small delay deviations should result in a rather uniform error distribution for the first refresh cycle, which is a good thing then, irrespective of the absolute error level.
Knowing the absolute luminance errors is useful when estimating the true luminous energy of a stimulus. But the errors also tell us something about the settling behavior on a slow time scale – considerably slower than what is implied by the 70 Hz low-pass filter.
Unfortunately, the LCD settling behavior depends quite a lot on the monitor settings, and it is impractical to measure and report the results for all the possible setting combinations. The monitor's factory settings should give a good starting point though, assuming that these are the settings the monitor has been optimized for. However, if the factory settings are too far off from any reasonable monitor calibration, more realistic settings should be used. One particular problem arises when colored scenes and color calibration comes into play. This is because the gains for the color channels (Red, Green, and Blue), along with the Contrast setting, define the maximal nominal range of operation, i.e., the maximal luminance per primary color. If these settings are not maxed out, the remaining range can be used internally by the monitor for overdrive. But even if overdrive is not active, the settling behavior of the LC cells still depends on the chosen operating range.
This makes it difficult to choose good settings for the measurements. On the one hand, the measurements (or derived parameters) should provide information about the maximal available operating range, allowing the user to make an educated choice about the extent the operating range should be possibly delimited for best performance given the application in mind. On the other hand, the measurements should reflect the monitor's performance under the most commonly used circumstances.
The Brightness setting, by the way, does not have a strong impact on the settling behavior.
- Side note: A higher Brightness setting might still have a small impact though as the temperature of the liquid crystal cells and, hence, the cell settling behavior somewhat depends on how powerful the backlight is. A change in temperature can either be caused by the additional heat coming from the light source or by the increased luminous energy passing through or being absorbed by the LC cells.
However, setting the Brightness to 100% usually makes even monitors with PWM backlight flicker-free, which is a necessity for running the measurement procedure successfully. Moreover, a higher brightness also results in a better S/N ratio for the photodiode measurements.
- Side note: To some extent, Contrast and Brightness settings are seemingly interchangeable as both have a major effect on the white luminance but only little or no effect on the black luminance. But the difference between the two actually lies in the effect they have on the black luminance. The Contrast setting indeed changes the black-to-white contrast by changing the white luminance but not the black luminance, whereas the Brightness setting does not change the black-to-white contrast as it affects the black and the white luminance to the same (relative) extent. Besides that, the Contrast setting usually affects the effective color accuracy as it changes the coding range for the internal color processing, whereas the Brightness setting does not as it just modulates the backlight. Therefore, it is normally useful to choose a rather high Contrast level, one that does not crush white levels and possibly leaves enough headroom for overdrive. The desired maximal luminance should then be adjusted via the Brightness setting.
Last but not least, because the settling behavior depends quite a bit on the temperature of the liquid crystals – after all, liquid crystals are fluids with varying viscosity over temperature –, it is crucial to let the monitor thoroughly warm up before doing the measurements (1 hr minimum).
The measurements have been taken with a photo diode PDA36A (Thorlabs), the gain of which was set to maximal 70 dB, resulting in a bandwidth of 12.5 kHz and a minimal rise/fall time (10%-90%) of about 30 µs. Sometimes, a gain of 60 dB was used for convenience as this was the gain that had to be used for the LightBoost measurements (because of the higher bandwidth it provides). The photodiode was placed at about 4.5 cm away from the screen surface at a straight angle. Ambient light was kept from the measurement by a rubber lens hood of 3 cm diameter which also limited the maximal incident angle to about ±20°.
The vertical extent of the measured screen area was not only limited by the rubber sleeve but also by the stimulus being just a small horizontal stripe covering 5% of the screen height. Note that the LC cells are updated sequentially from the top of the screen to the bottom, which results in different delays for the luminance curves depending on the vertical measurement position. By limiting the measurement to only 5% of the full vertical screen size, the smear effect introduced by averaging over differently delayed luminance signals becomes close to irrelevant. For example, for a refresh frequency of 120 Hz the screen is updated within around 8 ms, so if the true luminance would change instantly, the measured rise time would be 5%·8 ms = 0.4 ms which is negligible here. Other than for the smear effect, the delay of the luminance curve is of no further interest here. Nevertheless, the photodiode was placed at half of the screen height.
The photodiode was connected to a PC oscilloscope PicoScope 4224 (Pico Technology). At the hardware level, a 1 MHz sampling rate was used, but in the recording software the signal was strongly low-pass filtered (Gaussian with a -3dB frequency of 70 Hz) and down-sampled to 5 kHz (except for the trigger analysis, see below).
For the signal analysis it is important to know exactly at which point in time the luminance starts changing. This time depends on the screen location the signal was measured at. Even if there was a hardware signal available, like VSYNC, which indicated when the computer sent out a new image update and which could be used for triggering the oscilloscope at the hardware level, the update time of the measured area would still depend on the vertical position of the photodiode and on delays caused by the monitor electronics. So we better infer the trigger point directly from the signal. The beginning of the refresh cycle was defined by the very beginning of the falling edge of a single luminance trigger pulse (in a black-white-black frame sequence). More precisely, the maximum derivative of the falling edge was determined and the knee point of the falling edge was defined by where the derivative was just a 6th of this maximal derivative. This was done in the same way for two trigger pulses – one before and one after each frame sequence comprising a luminance step. Measuring two trigger pulses per sequence allowed to determine the exact refresh rate in terms of oscilloscope samples and thus provided also the means for checking the timing consistency. In order to maintain the same conditions for each trigger frame, trigger frames were preceded by at least two black frames and were followed by one black frame. Moreover, the trigger frame was a light gray (pixel value 240) instead of fully white (255) in order to avoid driving the LC cell close to or into saturation, which otherwise could have caused the cell to take a while before any luminance change could be observed when switching back to black. This would result in misleading values for the refresh timing.
In addition to the trigger pulses, the frame sequence for each particular luminance step consisted of 11 frames with the pre-step pixel value, and 10 frames with the post-step pixel value. The average luminance over the two last frames of the pre- and post-step sequence defined the pre- and post-step luminance respectively. Directly measuring the settled pre- and post-step luminance levels allowed to get rid of rather high offset fluctuations of the photodiode (not clear whether these fluctuations were present all the time and where they came from – the photodiode, the amplifier, or the power supply).
The complete luminance step matrix (stepping from/to 9 luminance levels) was measured systematically, cycling through the "from" levels in the outer loop and through the "to" levels in the inner loop. This matrix measurement was repeated for N=5 times in order to assess a measure of variation (shown as standard deviation by the thin green bars in the plots).