Chapter 4 Adaptive LC/BL Feedback Control
4.5 Results
Five images in Fig. 4-9 were scrolled with a speed of one image per second as a test video to verify this CBU reduction method. Fig. 4-13 shows the ΔEsum variation with time in a frame sequence. During the first three frames of each image, the ΔEsum is confirmed to be decreased and stable by the feedback control of ΔEsum minimization. Thus, the feedback control technique can effectively determine optimized BL values. In addition, the ΔEsum of proposed DRGB method is the lowest in comparison with those of conventional RGB and C/YRGB ones. This average CBU suppression ratio is around 70% of conventional one as shown in Fig. 4-14. According to the evaluation of observations, the CBU artifact as expected is not noticeable.
Time (second)
Fig. 4-13 The variation of color difference ΔEsum with a scrolled speed of one image per second for these test images.
Fig. 4-14 The ΔEsum comparison of test images with conventional, c/y, and dominated RGB methods and the ratio of ΔEsum of DRGB to conventional one.
Fig. 4-15 shows the typical CBU and modified images by D-field with the minimum ΔEsum. In the circle marks, it is obvious that the CBU artifact was greatly reduced in comparison with that of conventional RGB 3-field sequence. The color separation was greatly improved at the edges of Tiffany’s face and robot’s body.
Similarly, the bright and dark sides of mountains in Airplane presented unnoticeable color breakup. The wing of nose in Baboon and the brim of hat in Lena also showed the reduced CBU significantly. The experiment results of perceived images were agreed with the observation. With the current platform of BL determination on D-field, the optimized value of each frame is obtained after two frame time. In order to determine this value during the same frame time, the parallel architecture of feedback loop can be designed to speed up the processing steps. A real-time CBU reduction becomes realizable for fast moving images.
Fig. 4-15 The CBU images with (a) conventional RGB and (b) adaptive DRGB fields.
Tiffany
Space robot
Airplane
Baboon
(a) typical CBU
(b) D-field CBU Lena
Tiffany
Space robot
Airplane
Baboon
(a) typical CBU
(b) D-field CBU Lena
4.6 Summary
An adaptive feedback control for gray level rearrangement of LC/BL signals to reduce CBU artifact was demonstrated on a 32” FSC-LCD. The color backlight of the D-field with the minimum color difference between CBU images and original ones was proposed to suppress the CBU artifact effectively. According to the image content, the adapted color backlight can concentrate the light intensity on the dominated field and minimize the CBU effect. The 2x4 sampling period of image and 3-bit gray levels of color backlight were applied for the simplification of hardware implementation.
The field frequency of 240 Hz, the brightness of 400 nits, and the total power consumption of 50W were achieved. Most importantly, the 4 fields per frame have been fulfilled compared to the other techniques for reducing CBU effect. With rearranging LC/BL signals dynamically and the proposed feedback algorithm of optimized color backlight, our results successfully demonstrate that the proposed method is a practical way to suppress the CBU in field sequence color LCD applications.
Chapter 5
Local Adaptation and Boundary Issue
So far, the color of D-field is modulated on the whole backlight, or called the global adaptation. To contain more intensity of brightness in D-field and further reduce the CBU artifact, local adaptation can be applied along horizontal and vertical segments (2D-dimming) for locally optimizing CBU on corresponding areas of image [55]. The proper light distribution in a segment is required to achieve whole backlight profile as uniform as possible when each segment switched on the same level [87][88].
On the other hand, the boundaries should not be visible while 2D dimming. In this thesis, we investigated the effects of optical profiles on the perceived image quality.
The threshold of boundary-free image was found according to human contrast sensitivity function. Simultaneously, the proper light profile in each segment was derived for the design of large-sized FSC-LCDs (the third part of thesis architecture in Fig. 5-1).
mobile laptop TV
(Basic CBU suppression) (Further CBU suppression)
(Design Recommendation)
consistent color (global arrangement)
dominated color (global adaptation)
2D-dimming color (local adaptation)
mobile laptop TV
(Basic CBU suppression) (Further CBU suppression)
(Design Recommendation)
consistent color (global arrangement)
dominated color (global adaptation)
2D-dimming color (local adaptation)
Fig. 5-1 The thesis architecture.
5.1 Contrast sensitivity function considered
The human contrast sensitivity function is analyzed in order to determine a condition for boundary-free perception in 2D-dimming backlight. From the perceptive experiment, the contrast sensitivity function determines the threshold amplitude value of sinusoidal variation to indistinguishable perception [89][90]. At one particular frequency (cycles/degree), human eyes have different ability to perceive the light intensity variation. The sensitivity is the inversion of the contrast in Fig. 5-2[89]. For an example, at 1 cycle/degree, human can not sense less than 1/100 variation. The most sensitive frequency is 5 cycles/degree, equivalent to 20 pixels on TV application according to the viewing distance and pixel size in Tab. 5-1[90]. One segment with realistic cost level on 2D-dimming backlight covers more than 20 x 20 pixels on LC panel. Therefore, we will focus on lower frequency, as shown in left part of Fig.
5-2(a), to find the relation between contrast sensitivity and threshold of boundary-free profile.
(a) (b)
Fig. 5-2 (a) Contrast sensitivity function and (b)with experimental values
Tab. 5-1 The relation between pixel number and frequency in LCD TV and laptop
The gradient of backlight image was analyzed to evaluate the boundary perception. We start by considering the boundary of just not noticeable that is drawn in Fig. 5-2(a).
where P(x) is the sinusoidal variation of luminance at this boundary in Fig. 5-2(a), a(f) is the maximum amplitude at the boundary of just not noticeable, f is frequency (cycle/degree), the distance x (unit: degree) away from origin, and an off-set value of 1 for P(x) as a positive number. For an example, when f is 1 cycle/degree, a(f) is 1/100 of background luminance (LB) according to contrast sensitive function shown in Fig.
5-2(b). When the background luminance is higher than 10fL (=34.3 nits), the contrast is independent of the background luminance [91]. Then, the first derivative of P(x) is,
) When the frequency is lower than 5 cycles/degree, the contrast sensitivity increases and the contrast decreases with the spatial frequency. The decrease in the contrast implies the decrease in the amplitude a(f) of luminance variation of just not noticeable. According to Fig. 5-2(b), a(f) is inversely proportional to the spatial
8
Distance between two points on screen with 0.2 degrees (5 cycle/degree) (mm)
Distance between two points on screen with 0.2 degrees (5 cycle/degree) (mm)
frequency. Therefore, the frequency factors of a(f) and 2πf are cancelled out in P’(x).
In other words, the amplitude for the boundary of just not noticeable in Eq. 5-2 will be independent of frequency with constant K of πLB/100 (unit: nits/degree), the maximum slope of undistinguished profile.
5.2 Verification using Lorentz distribution
The effect of backlight profiles on perceived image quality was investigated by MatlabTM to simulate 2D-dimming backlight image. A 256 x 256 resolution of an original image was divided into 8 x 8 segments as an example. The quadratic Lorentz distribution L(r) in each segment was chosen for the best match with the measured profile of each segment56. This distribution L(r) can be expressed as:
,
where Δr is the distance of a pixel to origin, the center of each segment, and σ determines the width of profile. The distribution decays with the distance away from the center of each segment. Moreover, the parameter of σ is used to modify the distribution L(r) as the light profile of each segment. A larger σ induces a wider segment profile, resulting in more overlaps between segments. In full-on condition, the peak values in each segment were set to equal and whole backlight image was obtained by the superimposition of segment profiles. Because the Lorentz distribution has a long tail, there are overlaps with neighboring segments. The result, shown in Fig.
5-3(a), is close to a uniform profile while σ is equal to a factor of 1.2 wider than the width of segment. In 2D-dimming condition, each peak value was determined by the maximum LC signal of 32 x 32-pixel in each segment. A picture of camera man and its backlight image are shown in Figs. 5-3(b) and (c). Boundaries in the backlight
image are close to indistinguishable. Fig. 5-3(d) shows the gradient of normalized backlight image in Fig. 5-3(c). We find the gradient values on the boundary between segments are lower than π/100.
(a) (b)
(b)
(c) (d)
Fig. 5-3 (a) Full-on backlight, (b) original of camera man, (c) 2D-dimming backlight, and (d) the gradient of 2D-dimming backlight.
Five images, Ninja, Snow, Clockbuilding, Game, and City with different spatial scales were further investigated. Fig. 5-4 shows these original images, the results of 2D-dimming backlight images and their gradient images. The contour values of gradient are presented with 100x amplification. In images of Ninja and Snow, the spatial scale is coarse, resulting in blurring backlight images. Not only on the boundary between the segments but also on the edge of objects in these two images,
Image gradmag
Image Image
the gradient values, for the most part, are lower than 4/100 (4 on contour). However, in Clockbuilding and Game, the images with intermediate spatial scale, the edges of objects in these two images are more obvious. Even though the higher gradient values are presented, the boundary between segments is still undistinguished. We also tested the effect of backlight profile for the fine detail image. To show the detail, the segments corresponding to the buildings in City were switched on. With the superposition of profiles, the gradient values are lower in the bottom part of image.
High gradient values appear on the edge between the dark sky and building in City, resulting from the switch-off segments of dark sky. Therefore, this worst case between brightest and darkest segments needs to be further investigated for boundary-free backlight image.
Ninja(a) Ninja (b) Ninja (c)
Snow(a) Snow(b) Snow(c)
Clockbuilding(a) Clockbuilding(b) Clockbuilding(c)
Game(a) Game(b) Game(c)
City(a) City(b) City(c) Fig. 5-4 The selected images over different spatial scale. (a) Original image, (b) 2D-dimming backlight, and (c) the contour of gradient image (100x)
5.3 Modified profile for 2D-dimming backlight
The proper light distribution in a segment is required to achieve whole backlight profile as uniform as possible when each segment switched on the same level. On the other hand, the boundaries should not be visible while 2D dimming. Moreover, the distributions should be substantially local for maximizing contrast. Therefore, we used the contrast ratio, the maximum to minimum value of backlight images, to examine the enhancement factor. Fig. 5-5 shows the results as the segment size of 0.25 to 4 degrees. The wider size of segment lowers the contrast ratio of these five images. The average factor of these images is close to 5 at the size of 1 degree.
Fig. 5-5 The backlight contrast ratio per images as a function of segment size.
The solid line corresponds to the contrast ratio averaged over all the images.
0.1 1 10
2 4 6 8 10 12 14 16 18
Segment size (degrees)
Backlight contrast ratio
Ninja Snow Clockbuilding Game City Average
Fig. 5-6 The percentage of pixel numbers below the threshold per gradient images as a function of segment size. The solid line corresponds to the average over all the images.
Further considering the boundary perception, the percentage of pixel numbers below the threshold in a gradient image is used to evaluate the visibility of boundaries.
The average value over 60% is obtained while the size of segment is wider than 1 degree as shown in Fig. 5-6. A simple recommendation is to utilize the size of 2 degrees for the balance between contrast enhancement and boundary perception. At this size, the gradient is close to 100% below the boundary-free criterion and the average contrast ratio is slightly lowered from 5 to 3.5.
To maximize the contrast ratio and minimize the boundary artifact, a spatial filter can be utilized in 2D-dimming backlight. The blurred image is simulated by convolving this filter with the backlight image. An average filter, as a low pass filter in frequency domain, is a candidate to blur the backlight image. The criterion of the size of segment to maximize the contrast ratio can be modified by the size of filter. In
0.1 1 10
20 30 40 50 60 70 80 90 100
Below theshold ratio (%)
Segment size (degrees)
Ninja Snow Clockbuilding Game City Average
a real system, to modify the size of spatial filter implies to adjust the scattering abilities of diffusers in a backlight module. Fig. 5-7 shows the percentage below the threshold and the contrast ratio with filter sizes of 0.25, 0.5, and 1 degree. Using spatial filters at the size of segment of smaller than 0.5 degree, the below threshold ratio is improved from 40% to 50% and the boundary artifact is alleviated. If this artifact is further suppressed, the maximum contrast of about 20 is achieved at the size of segment about 0.125 degree.
Fig. 5-7 The backlight contrast ratio and the percentage of pixel numbers below the threshold of gradient images as a function of segment and spatial filter size.
The lines correspond to the average over 5 images.
0.1 0.2 0.3 0.4 0.5 0.6 35
40 45 50 55 60 65
0.1 0.2 0.3 0.4 0.5 0.6 5
10 15 20 25 30
without spatial filter(SF) SF size of 0.25 degree SF size of 0.50 degree SF size of 1.00 degree
Below theshold ratio (%) Backlight contrast ratio
Segment size (degrees)
Fig. 5-8 shows the relation between the segment size and number of screens, determined by viewing distances corresponding to factors of screen diagonal. At a viewing distance of a factor of 3 longer than screen diagonal, the segment size of 2 degrees results in about 10x10 as the recommended segment number. For a closer viewing distance, more segment numbers are needed. This function between viewing distances and segment numbers is useful to the screen sizes from monitor to TV-sized application.
Fig. 5-8 The segment numbers of screen as a function of segment size. The lines correspond to the viewing distances of factors of 1.5 to 6 longer than screen diagonal.
1 2 3 4 5
1 10 100
Segment numbers of screen width
Segment size (degrees)
1.5X 3.0X 4.5X 6.0X
5.4 Proper light distribution formula
The proper light distribution in a segment is required to achieve whole backlight profile as uniform as possible when each segment switched on the same level. On the other hand, the boundaries should not be visible while 2D dimming. Moreover, the distributions should be substantially local for maximizing contrast [92][93]. So far, the Lorentz with σ=1.2 could be a good choice. In order to find whether other light distributions are suited for 2D-dimming, a simple model is analyzed as follows.
First, the origins are set in the center of each segment, and the pitch between two origins is set to 1. We choose three points located in origin (A), middle point of segment edge (B), and crossing point of segments(C), respectively, as shown Fig. 5-9.
These three points are significant for the light distribution. At point A, overlap light is contributed by four closest light sources and itself. For B, it is contributed by two closest light sources and other four sources. Similarly, four closest light sources contribute to the point C. Tab. 5-2 shows the relation between distance and number of contributed sources. Contributions from the distances larger than square root of 1.25 are cut out. Later we will show that this cut-off is justified.
Fig. 5-9 Three points in the model for light source contribution analysis
Tab. 5-2 The relation between distance and contribution of light sources
In uniform condition, these three points should have the same intensity, which can be expressed as:
)
where I(r) is a function of light profile. The square of distance is approximately inversion proportional to contributed number of sources in the first item on point B and C (2/0.25: 4/0.5). The possible form I(r) is assumed to be
A
,
where r is distance from origin, k determines the width of profile, and a is the power factor. One is added in denominator to avoid infinity at origin and normalize the contribution to one. Summation of the intensity differences between each two points is used to optimize a and k. In Fig. 5-10(a), the intensities of these three points are calculated with an extended 7x7 segments, whose cut-off distance about 3.5. While k around 0.8 to 1.2 and a around 2.5 to 2.8, these intensities are approximated. When the segments is increased to 15x15 or higher, a and k are converged on 2.2 and 1.0 as shown in Fig. 5-10(b). Convergence is achieved when the number of segments is larger than 15 x 15.
Fig. 5-10 Summation of differences and parameters in (a) 7x7 array (b) 15x15 and 31x31 array
This form of light distribution has a long tail, thus it degrades the contrast enhancement and worsens the localized ability. In order to improve contrast of backlight image, we add another power factor b as
,
to sharpen the profile of segment. In one condition of b equal to 2, the intensity difference is converged when segment array is 7x7 or higher. The parameter (a, k) are (2.3, 1.4) as shown in Fig. 5-11(a). We further increase b to 3, and modify again the (a, k) values to (2.2, 1.5) as shown in Fig. 5-11(b). With higher b such as 4, the gradient of backlight image will exceed the boundary-free criterion. Surprisingly, the parameter (a, b, k) equal to (2.2, 3, 1.5) so matches the Gaussian distribution with width factor equal to pitch of segment as shown in Eq. 5-7 and Fig. 5-12. The full-on and 2D-dimming backlight images of camera man are shown in Fig. 5-13. Gaussian distribution has the shortest tail of these three profiles, thus the contrast enhancement is the highest. The boundary issue can be avoided because the gradient of image is less than threshold. Moreover, the most uniform backlight image of these three ones can be achieved.
Fig. 5-11 Summation of differences and parameters with sharp factor (a) b=2 (b) b=3
Fig. 5-12 Similarity between a, b, k model , Gaussian, and Lorentz distributions
Fig. 5-13 (a) 2D-dimming (b) full-on backlight image of Gaussian distribution
-6 -4 -2 0 2 4 6
0.0 0.2 0.4 0.6 0.8 1.0
Position/ pitch
a,b,k model Gaussian Lorentz
Image Image
5.5 Summary
The human contrast sensitivity function was analyzed to determine the threshold of boundary-free perception in 2D-dimming backlight. The gradient of backlight image was used to evaluate the visibility of boundaries. From a perception study, it was found that the gradient of backlight image should be lower π/100 of background luminance. This value appears to be the threshold for the visibility of boundaries. The Lorentz distribution of backlight profiles was examined on images over different spatial scales, ranging from coarse to fine detail. We focus on perceived image quality of dimmable backlight with an economical number of segments for 2D-dimming backlight. The effects of size and profile of segment were studied based on human visual properties. Considering the boundary perception and contrast enhancement, the size of 2 degrees was found to the recommended profile of each segment. The gradient is close to 100% below the boundary-free criterion and the average backlight contrast is 3.5. Simultaneously, the proper light profile in each segment was derived for high uniformity of three significant points. Considering a better localized ability and high contrast enhancement, the Gaussian function was found to be the most suitable profile of each segment in 2D-dimming backlight.
Chapter 6
Conclusion and Future Work
6.1 Conclusion
The field sequential color liquid crystal display (FSC-LCD) has emerged as a new branch of LCD application. Sequential driving LED backlight represents a potential technological breakthrough in terms of optical efficiency. Because of its needless color filter and low material cost, the FSC-LCD has become the key technology for reducing the power dissipation and the resource consumption.
FSC-LCDs rapidly flash the primaries time-sequentially such that the colors are mixed by means of temporal integration in the eye. The lacking sub-pixels and color filters result in high transitivity and large aperture ratio. The primary chromaticities are determined solely by the LEDs which enable wider gamut and scalable number of primaries. Furthermore, the impulse driving of the backlight ensures high moving image quality.
However, color breakup (CBU) is the most disturbing artifact, which occurs in FSC-LCDs. The CBU reveals itself in the appearance of multiple color images of stationary object during saccadic eye motion, or along the edges of moving objects when tracking the objects with the eye. Although increasing the frame to several thousand Hz can completely eliminate CBU [94], it is highly unlikely that affordable panels with frame rates higher than 240 Hz will be widely available in the foreseeable
However, color breakup (CBU) is the most disturbing artifact, which occurs in FSC-LCDs. The CBU reveals itself in the appearance of multiple color images of stationary object during saccadic eye motion, or along the edges of moving objects when tracking the objects with the eye. Although increasing the frame to several thousand Hz can completely eliminate CBU [94], it is highly unlikely that affordable panels with frame rates higher than 240 Hz will be widely available in the foreseeable