TODAY, notebook computers, palm-size PCs, personal data

Transcription

1 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 12, NO. 8, AUGUST DLS: Dynamic Backlight Luminance Scaling of Liquid Crystal Display Naehyuck Chang, Member, IEEE, Inseok Choi, and Hojun Shim, Student Member, IEEE Abstract Backlight systems dominate the power requirements of battery-operated hand-held devices with color thin-film transistor (TFT), liquid crystal displays (LCDs). We introduce dynamic luminance scaling of the backlight with appropriate image compensation. Dynamic backlight luminance scaling (DLS) keeps the perceived intensity or contrast of the image as close as possible to the original while achieving significant power reduction. DLS compromises quality of image between power consumption, which fulfills a large variety of user preferences in power-aware multimedia applications. DLS saves 20% to 80% of power consumption of the backlight systems while keeping a reasonable amount of image quality degradation. Index Terms Energy management, image processing, liquid crystal displays, power measurement. I. INTRODUCTION TODAY, notebook computers, palm-size PCs, personal data assistants (PDAs) and even cell phones are equipped with color thin-film transistor (TFT) liquid crystal displays (LCDs). A quality LCD system is now the default configuration for handheld embedded systems. An LCD panel does not illuminate itself and thus requires a light source. A transmissive LCD uses a backlight, which is one of the greediest consumers of power of all system components. A reflective LCD uses ambient light and a reflector instead of the backlight. However, reflective LCD is not suitable for quality displays, and complementary use of the ambient light and the backlight, named transflective LCD, is used for small hand-held devices. When the backlight is turned off, a transmissive LCD displays nothing but black screen; even transflective screens are barely legible without the backlight. Table I summarizes three representative configurations of battery-operated systems with color TFT LCDs. The power consumption of the backlight ranges from 0.4 to 4 W, which is 20% to 30% of the total system power. MobileMark2002 reports that a notebook computer with Intel Centrino technology consumes 9.3 W power on average with 15 of luminance from the backlight [1]. The power consumption will be 13 W with 150 luminance from the backlight, which is a typical brightness [2]. We measured the power consumption of the backlight systems in palm-size PCs and PDAs and verified that they consume from 20% to 30% of the total power. The ratio becomes larger if the devices adopt recently developed advanced power Manuscript received October 8, 2003; revised August 7, The authors are with the School of Computer Science and Engineering, Seoul National University, Seoul , Korea ( naehyuck@snu.ac.kr; thiskid@cselab.snu.ac.kr; hjshim@cselab.snu.ac.kr). Digital Object Identifier /TVLSI management techniques for the CPU, memory and other peripherals. We estimated that the proportion may easily increase to 40% with cycle-accurate energy measurement and system-wide energy simulation [3] [5]. Thanks to recent work on reducing the power consumption of CPUs, memory systems, and communication adapters, we can achieve the power requirements of those components by utilizing slack times without affecting the users. But there is no explicit slack time for the backlight for as long as the display device is turned on. Brute-force dimming or turning off of the backlight should be used as a last resort. Recent backlight reduction techniques attempt to maintain image fidelity when the backlight is dimmed. Initially, brightness and contrast shift were used [3] to pursue this goal, and this approach has provided the foundation for the backlight power management schemes introduced in this paper. It is well known that ambient luminance affects the visibility of LCD TFT panels, and, by taking account of this, backlight auto regulation [6] can also reduce the average energy requirement of the backlight. Simultaneous brightness and contrast scaling [7] further enhances image fidelity with a dim backlight, and thus, permits an additional reduction in backlight power. In an MPEG-1 video streaming application, this approach (including the necessary image processing) has been implemented [8] by adaptive middleware, without placing an extra burden on the streaming clients. Even more aggressive management of the backlight can be achieved by modification of the LCD panel to permit zoned backlighting [9]. In this paper, we introduce dynamic backlight luminance scaling (DLS) that adaptively dims the backlight with appropriate image compensation so that the user perceives similar levels of brightness and contrast with minor image distortion. We introduce the principles of DLS, image compensation, image enhancement methods, and context processing. We demonstrate DLS using an MPEG4 decoder and the personal basis profile of the Java 2 Micro Edition. II. LIQUID CRYSTAL DISPLAYS A. LCD Structure Fig. 1 illustrates the structure of a transmissive color TFT LCD. The LCD panel does not illuminate itself, but it filters light from the light source from the rear of the LCD panel. A reflective LCD uses ambient light as the light source of the LCD so as to eliminate the power-greedy backlight lamp. Instead, it has a reflector to reverse the direction of the ambient light from the front. However, reflective LCDs do not offer a quality display. Transflective LCDs are compromises between transmissive and /04$ IEEE

2 838 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 12, NO. 8, AUGUST 2004 TABLE I SYSTEM CONFIGURATION PER EACH TYPE Fig. 1. Rotation of liquid crystal depends on AC voltage applied to the electrode. The amount of rotation is linearly proportional to the tristimulus color values, A (C). reflective designs but are usually operated with the backlight on. The LCD is an additive color reproduction system. A color on an LCD can be matched by three primary colors, and. The qualitative match of a color is denoted by where,, and are the matching values for the color. The luminance of the backlight is denoted by where is the color matching value for the primary color of a backlight (typically white in TFT LCDs); is the control voltage of the high-voltage inverter; is the wavelength of the backlight; is the relative luminous efficiency; and represents the spectral energy distribution of the primary color. The luminance of a pixel whose color value is, is denoted by where is the product of the aperture ratio and the transmittance of the two polarizers, and the terms are the transmittance values of the three color filters. The aperture ratio is determined by the area of the black matrix. Note that, denotes normalized color values [10]. In this paper, we assume without (1) (2) Fig. 2. Characterizing the power-luminance relation of the backlight. loss of generality that the luminance of a pixel is a linear function of. Typical transmittance values of the color filter and the polarizer are 1/3 and 1/2, respectively. Modern LCDs have an aperture ratio of around 2/3. B. Characteristics of the Backlight The backlight must be white and colored backlight sources cannot be used. Currently cold cathode fluorescent lamp (CCFL) and a white light emitting diode (LED) are the most popular choices for TFT LCDs. So far, CCFLs are much more widely used. A white LED backlight circuit is trivial, CCFL requires a high-voltage inverter as shown in Fig. 1. Since a CCFL is not a planar light source, a diffuser is required to make it planar. We characterize a CCFL and a high-voltage inverter in terms of power and luminance by the use of an illuminometer (a Minolta LS-100) and a digital multimeter, as shown in Fig. 2. The LCD panel and the inverter are supplied as a pair; in this case LP064V1 from LG-Philips, which is 6 diagonal, has 18-bit colors and a screen resolution [11]. Fig. 3 shows that the luminance of the backlight is almost linearly proportional to the power consumption. The graph gives us quantitative information about the luminance-power relation as well. The inverter is supplied with 12 V as a power source and has an input terminal for luminance control. The luminance of the backlight

3 CHANG et al.: DLS: DYNAMIC BACKLIGHT LUMINANCE SCALING OF LIQUID CRYSTAL DISPLAY 839 Let us assume that an LCD panel, with vertical and horizontal resolutions of and, respectively, displays the th image,, where (4) We dim the backlight by changing the control voltage from to with. Then, is changed to such that. Now we recover the luminance of the image on the LCD panel by modifying to such that (5) Fig. 3. Steady-state luminance and power consumption of the CCFL backlight system versus control input. is controlled by applying voltage with a range of 3.3 to 0 V to the input terminal. As implemented, the inverter generates its maximum output when the control voltage is zero, (i.e., negative control input). To make things easier to understand, we will assume that the control input is positive, and the output voltage is proportional to the control voltage in Sections III and IV. III. DLS As shown in Section II, the luminance of the backlight is proportional to its power consumption. As we dim the backlight, the brightness of the image on the LCD panel is reduced, but we save power. The principle of DLS is to save power by backlight dimming while restoring the brightness of the image by appropriate image compensation [3], [12]. Brute-force backlight dimming is a traditional technique to save power consumption, but it reduces brightness, and thus, the display quality is degraded. By contrast, DLS does not sacrifice the overall brightness of the image but accommodates minor color distortions. To achieve the maximum power saving for a given color distortion limit, DLS dynamically scales the luminance of the backlight as the image on the LCD panel changes. For more aggressive power saving, DLS may also perform active color modification, which allows extreme backlight dimming. The principle of DLS is based on (2) and (3). We introduce four different image compensation algorithms, classified into three categories, in the following subsections. A. Brightness Compensation Brightness is the intensity of an image perceived by human eyes. Since the human eye and perception system are beyond the scope of this paper, we assume without loss of generality that the brightness is linearly proportional to the luminance of the LCD panel. This is because the LCD is a matched, additive color reproduction system, and we may assume that the normalized color value, where is the number of primary colors, is linearly proportional to the signal conceived by the cons in the human eye. From (3), we obtain If for all for and, then, (i.e., perfect brightness compensation). However, in practice, if such that, perfect compensation cannot be achieved. This happens as is significantly reduced, or large numbers of pixels populate bright areas. The amount of overflow,, is truncated and thus causes distortion. We use a histogram function,, to quantify the distortion ratio,, which is denoted by where is the color depth. The matrix is a matrix such that in an RGB-color space. We may derive from YUV and HSV color spaces accepting minor errors. In a YUV color space In an HSV color space,. Brightness compensation is implemented by a color transformation function of the form (6) (7) (8) (9) (10) The threshold is determined by the given value of and (7). In practice, only pixels in the brightest area are distorted after compensation while others keep their original colors and brightness. Fig. 4 illustrates an example of this brightness compensation and the accompanying distortion. The major computational overhead of brightness compensation is the construction of the transformation function and the transformation of each pixel color. Building the transformation

4 840 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 12, NO. 8, AUGUST 2004 Fig. 4. Example of brightness compensation. Fig. 5. Discrete histogram for a GUI application. function includes the construction of the histogram and determining a value for. Transformation can be performed either by multiplication and division operations, or table lookup, depending on the system architecture. The latter generally has less overhead as for a high-resolution screen. B. Image Enhancement Brightness compensation allows a significant degree of backlight dimming while keeping the distortion ratio reasonable, as long as the image has a continuous histogram which is not severely skewed to bright areas. Although all the histograms are discrete by definition, we express that a histogram is continuous if adjacent values are similar with each other, considering the original image before digitization of the color values. Discrete histograms generally make it difficult to determine a proper. As shown in Fig. 5, if we choose a value for which is lower than, then the distortion ratio significantly exceeds the limit because already contains more than 10% of the total pixels. Otherwise, the distortion ratio meets the requirement, but the amount of backlight dimming becomes insignificant. Unfortunately, most graphical user interface (GUI) components have discrete histograms. Image enhancement allows us to apply DLS for the images with discrete histograms where the brightest area dominates the image. In particular, we will use the techniques of histogram stretching and histogram equalization. Histogram stretching is an extension of brightness compensation, in that the histogram is stretched with respect to the low threshold as well as the high threshold. Histogram stretching is a mixture of contrast stretching and thresholding; it truncates data not only in the brightest areas but also in the darkest areas. It generally doubles the amount of backlight dimming that can be achieved, in comparison with brightness compensation. As shown in Fig. 5, we are able to set a value of which is more effective than, be-

5 CHANG et al.: DLS: DYNAMIC BACKLIGHT LUMINANCE SCALING OF LIQUID CRYSTAL DISPLAY 841 cause darker pixels are much less numerous than brighter pixels. The distortion ratio of histogram stretching is given by (11) where and are the lower and upper distortion ratios, such that. Histogram stretching implies contrast enhancement rather than recovery of brightness; the contrast enhancement is more desirable for GUI applications, where readability is the primary objective. The transformation function for histogram stretching is defined as follows (12) Histogram stretching outperforms brightness compensation but building the transformation function has twice the computational complexity. Sometimes, the colors of objects are not important, but we need maximum readability. Text-based screens often fall in this category. In such cases, we need to achieve more contrast to allow more backlight dimming. Histogram equalization is a useful technique for this purpose. The transformation for histogram equalization is given by (13) Image enhancement is also applicable to images with continuous histograms. Since histogram stretching is an extension of brightness compensation, there are similar distortion to the image. The majority of pixels preserve their original colors, and thus we can apply brightness compensation and histogram stretching for streaming images where interframe consistency must be considered. On the other hand, histogram equalization is not applicable to streaming images because in this case most pixels change their colors. Histogram equalization has a tendency to spread the histogram of the original image so that the levels of the histogram-equalized image span a wider range [13]. Thus, histogram equalization generally offers better readability than histogram stretching when the image has a discrete spectrum. The computational complexity for building the transformation function of histogram equalization is the same as that for brightness compensation. Since the transformation function is not a polynomial implementation, table lookup is desirable. C. Context Processing Histogram equalization generally outperforms histogram stretching in terms of readability for GUI applications, if the histogram is discrete. However, some minor colors may be merged into each other and are thus no longer distinguishable after histogram equalization. In the case of photographs, some minor colors may be merged into others or become more similar to each other. But, in the case of text, the number of pixels does not correlate with importance. So we never allow text to be merged into its background after histogram equalization. Context processing is a useful technique to prevent small foreground objects from having similar colors to their background after histogram equalization. If a foreground color and a background color become equal or similar after histogram equalization, context processing restretches their colors so that the distance between them in color space is a maximum. The transformation function is denoted by (14) where is the background color of an object whose color is. Context processing is a postprocessing step that can be applied after brightness compensation, histogram stretching or histogram equalization. Distortion ratio no longer has meaning if context processing is used. Context processing does not require the overhead of building a transformation function since it is not based on the histogram. However, transformation does require context information for the application. Context processing can be implemented by addition operations only. IV. IMPLEMENTATION A. Embedding the DLS Functionality DLS consists of two major operations: image compensation and backlight control. Both can be implemented by modifying: 1 ) the application program; 2) the windows system; or 3) the frame buffer device driver. 1) DLS in an Application: If we embed the DLS function in an application, we do not need a standard interface between the application and the DLS function. Also context information is freely available. However, each application will have its own structure which may be proprietary, so we may encounter porting issues. Implementing DLS at the application level may cause trouble in multitasking environment because the backlight is a shared resource. Thus, in general, an application that performs DLS must run in full-screen mode. This approach is suitable for applications that need direct access to the frame buffer or naturally contain essential information for building the transformation function. In addition, it helps to optimize implementation by eliminating unnecessary memory copy operations during transformation. 2) DLS in a Windows System: We need a unified DLS manager for applications running on a multitasking environment with its own GUI. Application programs call the functions defined in the windows system with the original color values, and the system performs the DLS function. The modified windows system requires new functionality in the application program interface (API) to support DLS function, such as initialization, enabling, disabling, recalculation of the histogram, etc. The main advantage of this approach is that applications can benefit from DLS without core modification, and backward compatibility is naturally maintained. When the DLS function is implemented in a windows system, some specific applications may have lower performance than when the DLS function is implemented at the application level.

6 842 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 12, NO. 8, AUGUST 2004 Fig. 6. Backlight luminance control circuit: (a) built-in common circuits for backlight dimming, (b) a built-in circuit for CCFL backlight, (c) a built-in circuit for a white LED backlight and (d) an add-on circuit to enhance the response time of a CCFL backlight. For instance, we can apply DLS to images in an compressed domain during MPEG decoding. If the compression ratio is 10, DLS implemented at the level of the windows system will need to perform 10 times more operations for color transformation. 3) DLS in a Device Driver: A frame buffer device driver maps the logical address of the frame buffer to its physical address. We can modify a frame buffer device driver to redirect the physical address to a DLS buffer. A daemon program polls the DLS buffer and performs the DLS function once the DLS buffer has changed. This approach does not require any code modification or the redesign of a windows system or an application. The overhead for rebuilding the transformation function is determined by update rate of the frame buffer. Since there is no context information coming from either the applications or the windows system, only a change to part of the frame buffer will cause rebuilding of the transformation function. Thresholding or other techniques can mitigate the overhead, but they may also cause noticeable visual artifacts. DLS in the driver also incurs additional frame buffer copy operations. B. DLS in an MPEG4 Decoder As an example of DLS implementation in an application layer, we implement the DLS function in an MPEG4 decoder running on Linux. In the case of this application, brightness compensation generally achieves enough backlight dimming, because most movie streams have continuous histograms. Some cartoon strips may have discrete histograms, but in general the brightest area does not dominate the screen. Histogram stretching would be more robust against different types of histogram, but needs more computational complexity. We perform brightness compensation after inverse Huffman coding and application of an inverse discrete cosine transform (IDCT). Since the decompressed images are in a YUV color space, we use luminance information to build the transformation function, using (9) rather than (8) and thus, reducing the computational complexity. We may perform the DLS function before IDCT and thus achieve further reduction of the computational complexity because the DC components of the DCT coefficients (the first coefficient in each 8 8 block) represent the luminance of the whole block. Using only the DC components results in minor errors in the histogram function. C. DLS in the Java 2 Micro Edition Personal Basis Profile As a demonstration of the DLS implementation in a windows system layer, we implement some lightweight abstract window toolkit (AWT) components 1 supporting the DLS function. The AWT components run on Microwindows [14]. We demonstrate DLS within a pseudomultiwindow environment running Java applications without modifying the windows system, i.e., Microwindows. The implementation includes coding the lightweight AWT components, since the personal basis profile (PBP) does not include them. The lightweight AWT components are transparent to developers and users, and compatible with existing Java applications. Backlight luminance control is performed by a native code. The lightweight AWT components maintain the context information of the objects. Histogram stretching, histogram equalization, and context processing are available under the Java runtime environment for applications with discrete histograms or displayed texts. The lightweight AWT components basically inherit the java.awt.component. The top-level container generates a DL- SEvent when the color information has been updated. The color information of the components is changed by recursive calls of the brighter method from the top to the leaf containers. D. Controlling the Backlight CCFL is a popular backlight for transmissive LCDs, although, recently white LED backlights have found increased applications in small hand-held devices. Basically, most devices are equipped with luminance-controllable backlight systems as shown in Fig. 6. The luminance control is used to extend battery life or to adjust to the luminance to accommodate changes in ambient light, and the users preferences. For a CCFL backlight, items (a) and (c) are installed while (a) and (b) are installed for a white LED backlight. Simply updating the values in the backlight luminance register changes the luminance of the backlight. Because a white LED backlight is responsive, open-loop luminance control is enough for dynamic luminance 1 The Java class library provides a GUI toolkit called the AWT. Typical GUI components include buttons, scrollbars, and text fields. Components allow interactions with the program and provide visual feedback to users. The GUI components can be created in the form of heavyweight components or lightweight components. Among them, the lightweight components are rendered purely in Java, without the use of platform-dependant peers.

7 CHANG et al.: DLS: DYNAMIC BACKLIGHT LUMINANCE SCALING OF LIQUID CRYSTAL DISPLAY 843 Fig. 7. Feedback control of the CCFL backlight (a video stream). scaling. On the other hand, the step response of the CCFL is an immediate step change plus a first-order exponential response whose time constant is well over tens of seconds. This is too slow to allow CCFL to be used for dynamic luminance scaling. Thus, we must apply feedback control to enhance the response. Simple PID control is enough for this purpose (15) Fig. 8. DLS test platform. where,, and are the control command input, the control voltage of the backlight inverter, and the error at th control loop, respectively. The photo diode generates a fairly linear response to the backlight luminance, and is a function that maps the voltage of the photo diode to the inverter control voltage. We tuned the PID coefficients to the following values:, and with. These settings achieved a 20 ms 10% to 90% rise time (2000 ). Fig. 7 shows that the luminance of the backlight closely follows the command input. Fig. 8 shows the prototype implementation of an LCD system supporting DLS. We built the prototype from a Linux PC, an LCD controller card and a 6 color TFT LCD with a CCFL backlight. As shown in Fig. 8, the LCD panel is mounted on a circuit board together with a simple microcontroller for feedback control. 2 demonstrated early results at the ACM SIGDA University Booth at Design Automation Conference 2002, and at the International Symposium on Low-Power Electronics Design As an alternative to prototype demonstration, we can prove DLS and the effect of image compensation for a dimmed backlight by the use of simulated brightness, following (3). Since the brightness of the image on the LCD panel is the product of the luminance of the image and the luminance of the backlight, we present the perceived intensity (brightness) of the images by scaling the color values by the relative luminance of the backlight. The reference luminance of the backlight is 100 cd m, and we can visualize an image with a backlight luminance of 50 cd m, by drawing a simulated image whose color values are half those of the original. V. EXPERIMENTAL RESULTS A. Presentation of Compensated Images Since we have a working prototype LCD system with DLS, performance analysis has been done by measurement. We 2 A PIC microcontroller and an embedded AD converter achieve a 1-ms control loop for testing purpose; Linux is not capable of 1-ms-period tasks. B. MPEG4 Decoder The performance index of DLS is the amount of backlight dimming achieved. In this paper, we use the relative luminance of the backlight and absolute power reduction as performance indices. In general, relative luminance is equal to the amount of power saving unless the backlight luminance fluctuates rapidly.

8 844 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 12, NO. 8, AUGUST 2004 Fig. 9. Power saving with DLS [distortion ratio (%), a video stream]. Fig. 10. Power saving with DLS [distortion ratio (%), a cartoon stream]. TABLE II POWER REDUCTION BY DLS Fig. 9 illustrates a luminance-scaling sequence for a movie stream, with 1% to 5% distortion ratios. 3 The default luminance of the backlight is 100 cd m (the dashed line). Whenever is above the dashed line, we have achieved a power saving. Fig. 10 shows DLS for a cartoon stream. Since cartoons exhibits less continuous histograms, the luminance sequence includes many abrupt changes. Table II provides a clearer performance comparison. Another important performance index is quality degradation of the images. In this paper, we use distortion ratio for the index, butitisnotexactlyequaltotheactualimagequality.forthemovie stream, we measure acceptable image fidelity through a mean opinion score (MOS) test with a 5% distortion ratio. In practice, the distortion ratio will be determined by users preference. C. JAVA 2 Micro Edition In order to keep the colors of GUI components as close as possible under backlight scaling, we may perform brightness compensation or histogram stretching as we did for movie and cartoon streams. We would naturally prefer to preserve the original colors in GUI application under backlight dimming. However, sometimes users may allow aggressive color changes to achieve the longest possible battery life, as long as they can still recognize the screen contents. This behavior is not expected for movies or JPEG images. 3 From this section, we return to negative control input. Fig. 11. DLS experiments on several movies. The photos on the left-hand side are screen shots of the original movies. The photos on the right-hand side are compensated screen shots. Power reduction/relative backlight luminance for each screen shot with D =0:03 was (a) 147 mw/87.5%; (b) 147 mw/87.5%; (c) 188 mw/82.0%; (d) 703 mw/42.1%; (e) 805 mw/35.9%; and (f) 278 mw/75.0%7

9 CHANG et al.: DLS: DYNAMIC BACKLIGHT LUMINANCE SCALING OF LIQUID CRYSTAL DISPLAY 845 Fig. 12. Image enhancement. (a) Original image. (b) Simulation of the original image with a dimmed backlight (50%). (c) Simulation of histogram stretching with a 10% threshold and a dimmed backlight (50%). (d) Simulation of histogram equalization and a dimmed backlight (50%). Fig. 14. Context processing. (a) Original image. (b) Simulation of histogram stretching with 10% threshold. (c) Simulation of context processing after histogram stretching with a dimmed backlight (50%). when the background and the foreground colors become similar to each other. Context processing uses information from the application to remedy the effects by changing the foreground color to the complement of the background, as shown in Fig. 14(c). Fig. 13. Image enhancement with the AWT components. (a) Original image. (b) Simulation of the original image with a dimmed backlight (50%). (c) Simulation of histogram stretching with a 10% threshold and a dimmed backlight (50%). (d) Simulation of histogram equalization and a dimmed backlight (50%). Fig. 12 shows DLS with image enhancement for a road map displayed on a global positioning system (GPS) guided navigation program. Fig. 12(b) illustrates that the original image [Fig. 12(a)] is barely recognizable under 50% luminance. As you can see in Fig. 12(c), the image quality is largely recovered after histogram stretching while 50% of the backlight power is saved. It is difficult to compare Fig. 12(c) and (d) even with the MOS test; nevertheless, the roads are definitely more distinguishable after histogram equalization [Fig. 12(d)]. Image enhancement generally recovers image quality under a dimmed backlight, but has a tendency to make minor colors merge within, or become indistinguishable from, other colors. We have already noted that we must be careful of text in GUI components. Fig. 13 has many characters and shows good readability after histogram stretching and histogram equalization where the luminance of the backlight is reduced to 50%. On the other hand, characters in Fig. 14(b) are barely recognizable after histogram equalization, because they occupy only small areas. Sometimes, image enhancement may result in poor readability VI. CONCLUSION Battery-operated hand-held systems are widely equipped with transmissive or transflective color TFT LCDs which require backlight. This is one of the dominant power consumers taking around 30% of the total system power. Based on the user s perception of image clarity, we introduce techniques for substantially reducing the power requirement backlight systems while providing minor or progressive color distortion. We call our approach DLS. We formulate a mathematical description of the DLS algorithms and demonstrate a working prototype. We compare three different implementation layers for the DLS functionality, and demonstrate two of them through implementation on an MPEG4 decoder and a Personal Basis Profile of Java 2 Micro Edition. Since the power consumption of a backlight is almost linearly proportional to its luminance, DLS reduces power consumption effectively. The DLS technique can be implemented in software without modifying the hardware for white LED backlight systems, or with minor hardware modifications for CCFL backlight systems. It does not require much computation, and the proper choice of an implementation layer reduces its complexity. DLS is waiting for final approval in MPEG-21 Standard, Part 7 Digital Item Adaptation. ACKNOWLEDGMENT The authors wish to thank Prof. N. Cho and his colleagues who helped implement the MPEG4 players supporting DLS; they thank H. S. Kim who implemented the PBP of Java 2 Micro Edition, and the Electronics and Telecommunications Research Institute (ETRI) who helped add the backlight luminance as a device capability in MPEG-21 digital item adaptation (DIA) for

10 846 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 12, NO. 8, AUGUST 2004 potential use of the DLS. They also thank the Institute of Computer Technology at Seoul National University and the Brain Korea 21 Project for providing some of the research facilities used in this study. REFERENCES [1] (2004) Intel Notebook Performance. Intel Co. Ltd. [Online]. Available: [2] (2002) LTN141XJ 14.1 XGA TFT LCD Preliminary Specification. Samsung Electronics Co. Ltd. [Online]. Available: [3] I. Choi, H. Shim, and N. Chang, Low-power color TFT LCD display for hand-held embedded systems, in Proc. Int. Symp. Low-Power Electronics and Design, Aug. 2002, pp [4] N. Chang, K. Kim, and H. G. Lee, Cycle-accurate energy consumption measurement and analysis: Case study of ARM7TDMI, IEEE Trans. VLSI Syst., vol. 10, pp , Apr [5] H. Shim, Y. Joo, Y. Choi, H. G. Lee, and N. Chang, Low-energy offchip SDRAM memory systems for embedded applications, ACM Trans. Embedded Computing Syst., vol. 2, no. 1, pp , Apr [6] F. Gatti, A. Acquaviva, L. Benini, and B. Ricco, Low power control techniques for TFT LCD displays, in Proc. Int. Conf. Compilers, Architecture, and Synthesis for Embedded Systems, Oct. 2002, pp [7] W.-C. Cheng, Y. Hou, and M. Pedram, Power minimization in a backlit TFT-LCD display by concurrent brightness and contrast scaling, in Proc. Design, Automation and Test in Europe, Feb. 2004, pp [8] S. Pasricha, S. Mohapatra, M. Luthra, N. Dutt, and N. Subramanian, Reducing backlight power consumption for streaming video applications on mobile handheld devices, in Proc. 1st. Workshop on Embedded Systems for Real-Time Multimedia, Oct. 2003, pp [9] J. Flinn and M. Satyanarayanan, Energy-aware adaptation for mobile applications, in Proc. Symp. Operating Systems Principles, 1999, pp [10] T. Tsukuda, TFT/LCD: Liquid-Crystal Displays Addressed by Thin-Film Transistors. New York: Taylor & Francis, [11] (1998) LP064V1 6.4 VGA TFT LCD Preliminary Specification. LG Philips LCD Co. Ltd. [Online]. Available: [12] I. Choi, H. S. Kim, H. Shin, and N. Chang, LPBP: Low-power basis profile of the Java 2 micro edition, in Proc. Int. Symp. Low-Power Electronics and Design, Aug. 2003, pp [13] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 3rd ed. Reading, MA: Addison-Wesley, [14] The microwindows project, G. Haerr. (2003). [Online]. Available: Naehyuck Chang (M 97) received the B.S., M.S., and Ph.D. degrees in control and instrumentation engineering, Seoul National University, Korea, in 1989, 1992, and 1996, respectively. Since 1997, he has been with the School of Computer Science and Engineering, Seoul National University, where he is currently an Associate Professor. His research interests include system-level low-power design and embedded systems design. Inseok Choi received the B.S. and M.S. degrees in computer science and engineering from Seoul National University, Korea, in 2002 and 2004, respectively. His graduate research includes low-power LCD display systems. Hojun Shim (S 01) received the B.S. degree from the School of Computer Science and Engineering, Seoul National University, Korea, in 2000, where he is currently pursuing the Ph.D. degree at the same university. His current research interests include working on low-power systems and embedded systems.