CONVENTIONAL image sensors, owing to a narrower

Transcription

1 3000 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 56, NO. 12, DECEMBER 2009 Dynamic-Range Widening in a CMOS Image Sensor Through Exposure Control Over a Dual-Photodiode Pixel Jung-Bum Chun, Hunjoon Jung, and Chong-Min Kyung, Fellow, IEEE Abstract In this paper, we propose a technique for automaticexposure control and synthesis for a wide-dynamic-range sensor based on dual-exposure method. Using the proposed technique, a sensor can adaptively adjust its exposure to widely varying illumination conditions and consequently accomplish an infinite dynamic range in theory. An artificial test bench consisting of a virtual illumination source and its imaging system is introduced to verify the performance of the proposed technique. Simulations show how a wider dynamic range can be achieved by the proposed technique. It is also shown that a VGA-resolution CMOS image sensor developed based on the proposed technique can support up to a 119-dB dynamic range. Index Terms CMOS image sensor (CIS), dual exposure, wide dynamic range (WDR). I. INTRODUCTION CONVENTIONAL image sensors, owing to a narrower dynamic range than the human eye, cannot properly reproduce a scene with a wide illumination range. Fig. 1 shows that, for a scene with a wide illumination range, a digital camera should adjust its exposure to either 1) the darker part or 2) the brighter part. This is a very serious limitation in surveillance and automotive applications where a wide dynamic range (WDR) capability is strongly required. In general, the dynamic range of an image sensor is defined as 20 log(max_lux/min_lux) where max_lux and min_lux refer to the maximum and minimum levels in the valid illumination range, respectively. Dynamic range can be increased by enhancing the signal-to-noise ratio (SNR) of a sensor pixel by noise suppression or signal enhancement. When the SNR is limited, WDR can be achieved by generating a logarithmic light-to-output signal transfer curve by employing unique pixel architectures [1], [2]. One disadvantage of these approaches is that they lead to low contrast and loss of details because they operate in continuous-time mode. Another disadvantage is the fixed pattern noise (FPN) generated from transistor mismatch, which is larger than that of a typical pixel structure. Manuscript received June 3, 2009; revised September 9, Current version published November 20, The review of this paper was arranged by Editor J. Tower. J.-B. Chun and C.-M. Kyung are with the Department of Electrical Engineering, Korea Advanced Institute of Science and Technology, Daejeon , Korea ( jbchun@vslab.kaist.ac.kr; kyung@ee.kaist.ac.kr). H. Jung is with Clairpixel Company, Ltd., Seoul , Korea ( henry@clairpixel.com). Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /TED Fig. 1. Conventional image sensors can only handle a much narrower range of illumination level than the human eye. (a) Overexposure is required for proper exposure of the indoor scene. (b) Underexposure is required for proper exposure of the outdoor scene. (c) A scene obtained by averaging the two images. S. Decker et al. proposed a stepped reset-gate technique to increase the dynamic range of a CMOS active pixel sensor array [3] where the transfer characteristic of a pixel relating input illumination to output voltage is intentionally changed from that of the standard pixel. The random noise and FPN, which are functions of illumination, also change. WDR can also be achieved by composing multiple images with different exposures from pixel structures with a built-in analog-to-digital converter (ADC) [4], [5]. Because they require many transistors per pixel, the fill factor and the input sensitivity of the sensor are low. The WDR implementation method described in [6] and [7] is based on dual exposure in a CMOS image sensor (CIS) where a WDR image is composed from two images with different exposure levels. An example of a WDR image based on dual exposure is shown in Fig. 1(c) which is the average of those in Fig. 1(a) and (b). The WDR implementation based on dual exposure is regarded as the most practical approach due to the following two reasons: First, it inherits the undegraded fill factor and the sensitivity of the conventional CIS pixel architecture, which enables the sensor to have a brighter and lower noise image in a fixed exposure time than other approaches [1] [5]. Second, since most scenes requiring the WDR function can be represented by two major illumination ranges, two separate exposures for the low and high illumination ranges are sufficient. In this paper, we propose a technique for automatic-exposure (AE) control and image synthesis for a WDR CIS based on dual exposure. To derive maximal information from scenes with a wide illumination range in real time, real-time exposure control and image synthesis must be considered when the sensor architecture is designed. The previous works focused only on sensor architectures that enable dual image acquisition without /$ IEEE

2 CHUN et al.: DYNAMIC-RANGE WIDENING IN CIS THROUGH EXPOSURE CONTROL OVER PIXEL 3001 addressing its corresponding exposure and synthesis methods [6], [7]. Simulations show that a theoretically infinite dynamic range can be achieved by the proposed technique. This paper also proposes a sensor architecture to produce synchronized dual-exposure images. This architecture has been implemented in hardware along with the proposed technique. The remainder of this paper is organized as follows. Problems of AE control and image synthesis for a dual-exposure system are defined in Section II, and a technique for AE control and image synthesis is proposed in Section III. The simulation results of the proposed technique are described in Section IV, and a WDR CIS sensor developed based on the proposed technique and architecture is presented in Section V. II. PROBLEM DEFINITION For the dual-exposure WDR sensor, we assume that a scene consists of a relatively dark and a bright part, which require long and short exposures, respectively, for proper acquisition of the scene data. The two resultant images generated by these two different exposure levels are called a long-exposure image (LEI) and a short-exposure image (SEI), respectively. In many conventional image sensors, a closed-loop control mechanism called proportional integral derivative (PID) control has been utilized [8]. The PID controller adjusts the exposure time of a sensor to let the average output intensity of the whole or partial sensor area keep track of an assigned reference level. In order to use the PID control in a dual-exposure WDR sensor, two PID controllers to the LEI and the SEI can be assigned, respectively, with their respective reference levels. In this case, the reference level for the LEI needs to be higher than that of the SEI such that the LEI represents the dark part and the SEI represents the bright part. As the WDR sensor, like conventional image sensors, should produce a single output image, the final output should be synthesized from the two images, i.e., the SEI and the LEI. When separate PID controllers are applied with two different exposure controls to maximize WDR performance and image quality, it is necessary to define an exposure parameter representing each image and to set the parameter at an appropriate target value. A default exposure parameter for a single PID control is the average of pixel values in an image. To run separate PID controllers, one can assign fixed target references to both controllers and let the average of each image data follow its target. The final image is then acquired through weighted summation of the pixel values of the two images in a fixed manner. However, since the average of each image, SEI and LEI, reflects both bright and dark parts, this method leads to the following two problems. 1) Long-exposure problem (LEP): The dark part is not properly exposed in the LEI when the size of overexposed area is large. 2) Short-exposure problem (SEP): The bright part is not properly exposed in the SEI when the size of underexposed area is large. For the LEI, it is desirable that the bright part be overexposed (saturated) so that the dark part will be properly exposed. Fig. 2. Problems occurring when a typical PID-based AE is applied to a dualexposure-based image sensor. (a) The exposure level of the dark part is affected by the size of the overexposed area even with the same average value. (b) The saturated area is still saturated even with a lower average value. However, because the exposure of the dark part is affected by the size of the bright part, a fixed reference level given to the PID controller as the average pixel value of the whole scene does not guarantee a proper exposure level of the dark part. In other words, when the reference value is fixed such that it is independent of the ratio between the size of the bright and dark parts, the exposure level of the dark part decreases as the size of the bright part increases, as shown in Fig. 2(a). On the other hand, in the SEI, the bright part in the SEI is supposed to be properly exposed even though the dark part is underexposed. However, because the exposure of the bright part is also affected by the size of the dark part, the bright part in the SEI is not always exposed properly even with the same reference. Although a lower reference value than that of the LEI is given to the SEI, the bright part saturated in the LEI may remain unchanged in the SEI, as shown in Fig. 2(b). III. AE CONTROL AND SYNTHESIS TECHNIQUE In this section, we propose an AE control to solve the problems outlined in the previous section. The suggested AE control is integrated in a WDR imaging system, shown in Fig. 3(a), consisting of two parts, an imaging part and an AE part. The imaging part generates two image data, i.e., I L for LEI and I S for SEI with their current exposure times denoted as T L and T S, respectively. Each time I L and I S are generated, the AE control is executed based on the given reference levels (R L and R S )to update T L and T S which are looped back to produce the next images. The AE control of long exposure is a modified PID control. The average of I L, which is I L, is calculated by the AV GL block, and a new reference level denoted as R L is generated

3 3002 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 56, NO. 12, DECEMBER 2009 Fig. 4. Pseudologarithmic exposure curve from (3) to enable the WDR capability of the final output. Fig. 3. Proposed AE control and system. (a) WDR imaging system consisting of an imaging part and an AE part. (b) The curve for ΔR L determination in the PRblock as a function of K, with the overexposure ratio defined as N S /N T (N S is the number of overexposed pixels and N T is the total number of pixels). (c) The exposure curve for the short-exposure imager as obtained by subtracting M T S /T L. by the PR block where R L is obtained by adding a corrective amount ΔR L to the given reference R L R L = R L +ΔR L. (1) ΔR L is given as a function F : K (0, 1) ΔR L (0,UL) represented by a curve in Fig. 3(b). K is the overexposure ratio given as N S /N T, where N S is the number of overexposed pixels, i.e., the pixels exceeding a certain level provided by the long-exposure imager, and N T is the total pixel count of the image. A larger K implies that the average of the pixel values is more affected by the overexposure area. From the curve in Fig. 3(b), ΔR L (> 0) is generated for K (α, β) where α, β, and UL (0 <α<β<1, 0 <UL<255) are userdefined variables to customize the curve. Unlike traditional PID controllers, the proposed PID controller adaptively changes its target reference based on the evaluated K. This prevents the average of the dark part from decreasing as the size of the bright part increases, thus providing a solution to the LEP. The rest of the AE operation for the LEI is the same as conventional PID operations in that the next T L is generated by multiplying corresponding PID coefficients to the error E and then adding the results to the current T L. In the AE operation of the short-exposure imager, another modified PID control is employed to solve the SEP by excluding the overexposed area of every LEI. Instead of providing the original SEI data I S, new image data I S generated by the SUB block are provided to the averaging block AV GS.Inthe SUB block, the exclusion of the saturation area is achieved by subtracting M T S /T L from the original image data where M is the maximum output level of a pixel by the imaging Fig. 5. Imaginary light source referred to as illumination scale is proposed for the simulation. For the mth entry, illumination indexes denoted as n m are assigned, and each entry is assumed to emit ( nm 1)-lux light. system, and thus, M T S /T L represents a value in the SEI corresponding to the onset of saturation in the LEI curve in Fig. 3(c) I S = { [ I S M TS T L ], if I S >M TS T L 0, otherwise. Owing to the subtraction in the SUB block, the newly generated image I S is darkened by M T S/T L but still contains valid data in the overexposed area of the corresponding LEI, which enables the AV GS block to reflect the bright part during averaging. The AV GS block sums all data from the SUB block and divides it by (N T N S ), thus generating the average I S as its output. Based on the average and given the reference value R S, PID control is performed to provide the next exposure time T S to the short-exposure imager. In the next step, two individual images are synthesized through weighted summation to produce the final output image where the weights determine code ranges assigned to SEI and LEI. We utilized the variable K in defining the weights to reflect the ratio of the dark and bright parts. Thus, the following can be given as the final step of synthesizing the LEI and SEI: (2) I O = I L (1 K)+I S K (3) where I O denotes the final output. By this equation, a wider code range is assigned to the SEI as K becomes larger. When K = 0, i.e., there is no overexposed area in the LEI, the SEI does not contribute to the final output. Since the larger area is likely to contain more information, our approach is a costefficient way to maximize the amount of information in the final output. This synthesis method is also shown in Fig. 4, which shows that a new pseudologarithmic exposure curve is generated from the equation.

4 CHUN et al.: DYNAMIC-RANGE WIDENING IN CIS THROUGH EXPOSURE CONTROL OVER PIXEL 3003 Fig. 6. Simulated images, where the pixel levels are plotted with respect to x-coordinate, from (a) the typical PID approach and (b) the proposed approach based on the illumination scale {(0,10), (23,245)} to show WDR capability. In the graphs, LEI, SEI, and synthesized output are represented as I L, I S,andI O, respectively. IV. SIMULATION To demonstrate the usefulness of the proposed AE control and measure its WDR capability, we set up an artificial test bench consisting of an illumination scale and a virtual imaging system. The illumination scale, which was inspired by a commercial gray scale target, is an imaginary light source proposed to express arbitrary illumination spectra. As shown in Fig. 5, it is an array of 32 rectangular light-emitting entries with different levels of illumination. The illumination level of the mth entry, denoted as its illumination index n m (0 n m 255), is( n m 1) lux. Hence, an entry can represent up to lux (= ). The mth entry of the illumination scale can be represented as a tuple (m, n m ). The illumination scale is defined as a set of 32 tuples {(0,n 0 ), (1,n 1 ),...,(31,n 31 )} in ascending order with respect to m or it can also be defined by less than 32 tuples in a way that the illumination indexes of undefined entries are obtained by incrementing by one, ascending from the defined entries, and the illumination indexes of the entries lower than the first defined entry are set to zero. For example, an illumination scale defined as {(m 1,n m1 ), (m 2,n m2 )} (0 m 1 <m 2 31) is an abbreviation of {(0, 0), (1, 0),..., (m 1 1, 0), (m 1,n m1 ), (m 1 +1,n m1 +1),...,(m 2 1,n m1 + m 2 m 1 1), (m 2,n m2 ), (m 2 + 1,n m2 +1),...,(31,n m2 m )}. For our test, we utilized the illumination scales defined by a set of one or two tuples because defining the illumination scale in this manner is appropriate to model a real environment which typically has one or two major light sources. The virtual imaging system, which is the other component of the test bench, is a software implementation of the imagers and the AE control shown in Fig. 3. Since the imagers are defined to map the illumination scale into a 32 1 pixel image, the final output is a 32 1 pixel image reproducing 32 illumination entries, each of which is reproduced into one pixel level. The imager is simply implemented as follows: I L (x, 1) =k L T L L(x + 1) I S (x, 1) =k S T S L(x + 1) (4) for 0 x 31, where L(m) represents the illumination level of the mth entry and k L and k S are the sensitivity coefficients for the long- and short-exposure imagers, respectively. The simulation was executed for arbitrarily generated illumination scales. The results of our method are compared with those of a conventional PID method. In our approach, the parameters are given as α = 0.2, β = 0.4, UL = 64, and R L = R S = 128, which are empirically determined so that they produce good results under most circumstances, while the reference levels of the PID controllers R L and R S are set to 128 and 72, respectively. The image synthesis is performed by afixedformula:i O = I L I S 0.3. Three simulation results with various dynamic ranges of illumination scales are shown in Figs. 6 8 where the pixel levels of the resultant images are plotted with respect to the x-coordinate, the entry number of the illumination scale. The illumination scale of the first example is {(0,10), (23,245)}, one of the widest illumination ranges of the proposed illumination scale. In the result of the conventional PID approach shown in Fig. 6(a), the dark part of its illumination scale is not properly exposed in its LEI although a lower reference level is given as its target value, which means that the reference level should have been lower than 72 for proper exposure of the dark part. However, there is no way of determining this value dynamically, which is the LEP introduced in Section II. On the other hand, the same part was better shown in Fig. 6(b) by the proposed AE control because R L is adjusted to a new value, which is 154, immediately after the AE operation commences. In the second example corresponding to a typical level of WDR capability, the illumination scale is set to {(0,30), (14,140)}. The bright part saturated in the LEI of Fig. 7(a) is not fully recovered in its corresponding SEI because the lowered reference value only further darkens the dark part. On the other hand, the information of the bright part is properly recovered in Fig. 7(b) in its SEI because the dark part was excluded when calculating its average value. This is our solution to SEP. The third experiment is for the case of a narrow dynamic range with the input being set to {(0,100)}. Fig. 8 shows that the final output of the proposed approach is more properly

5 3004 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 56, NO. 12, DECEMBER 2009 Fig. 7. Simulation images from (a) the typical PID approach and (b) the proposed approach based on the illumination scale {(0,30), (14,140)} to show mediumdynamic-range capability. Fig. 8. Simulation images from (a) the typical PID approach and (b) the proposed approach based on the illumination scale {(0,100)} to show narrow-dynamicrange capability. Fig. 9. Variation of the averaged pixel outputs of the images I L and I S with respect to the frame iteration count. (a) Conventional PID approach (b) Proposed approach. exposed than that of the conventional method. This is because the synthesis weight of each SEI is determined by the ratio of the overexposed region of the LEI, and as a result, the bright part contributes to the final output proportionally to its size. In Fig. 8(b), a single exposure for the LEI is enough to reproduce the scene because there is no overexposed part in the LEI, which prevents the SEI from contributing to the final output image. Thus, the output average is still 128, which means it is properly exposed. In the case of the conventional PID method, on the other hand, the fixed weight in its synthesis leads to underexposure in the final output image such that the average of the final image in Fig. 8(a) is only 90. Therefore, the adaptive synthesis method also can help the dual-exposure method to preserve the image quality even under an environment with a narrow illumination range. The dynamic range for the experiment is calculated from the formula 20 log(max_lux/min_lux), where max_lux and min_lux represent the maximum and minimum illumination levels, respectively. Since a noiseless sensor system is assumed in this experiment, the difference of the two lowest

6 CHUN et al.: DYNAMIC-RANGE WIDENING IN CIS THROUGH EXPOSURE CONTROL OVER PIXEL 3005 Fig. 10. Simplified schematic diagram of the implemented CMOS active pixel sensor comprising a pixel, CDS, and sense amplifier part. PD1 and PD2 are read out through separate paths after a common reset for CDS operation. illumination levels, which can be distinguished in the final output, substitutes min_lux. Hence, the WDR capabilities for the three simulations aforementioned are calculated as 122 db (= 20 log( /( ))), 77 db (= 20 log( /( ))), and 39 db (= 20 log( /( )), respectively. The proposed approach can support an infinite dynamic range depending on the definition of the illumination scale. However, in reality, the maximum dynamic range is bounded by physical limits or feasibility issues. Other test points of the evaluations are the convergence time and stability of the AE controls because the exposure should be stabilized at a proper level within a reasonable time without further fluctuation after the convergence. To show the convergence performance in terms of frame iteration count, the AE convergence graph for the case in Fig. 6 is shown in Fig. 9. The convergence time of the proposed method is still comparable to that of the conventional approach, because the major reference modification from R L to R L is finished within a few iterations from the start of the AE operation. Once a new reference is set, the average converges to the new reference as in the conventional PID control. V. S ENSOR IMPLEMENTATION We developed a CIS with WDR capability utilizing the dualexposure method based on the proposed AE control. Each pixel in the pixel array consists of photodiodes of two different sizes along with separate readout circuits. A concise schematic of the pixel and analog part is shown in Fig. 10, where PD1 represents the photodiode for LEI and PD2 for SEI. With other conditions set to be equal, the sensitivity, as well as the SNR of the pixel, is proportional to the size of its photodiode. Since we set the size of the long-exposure photodiode to

7 3006 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 56, NO. 12, DECEMBER 2009 Fig. 11. Developed WDR-capable CIS. (a) Block diagram. (b) Layout. (c) Sensor image (CLCC package). be four times larger than that of the short-exposure photodiode, theoretically, for coefficients k L and k S in (4), k L = 4k S. Increasing the size of a photodiode provides a means to improve its dynamic range by enhancing SNR and to effectively reduce the noise level. Both photodiodes in a pixel are shaped and located as shown in Fig. 11(a) to minimize the negative effect on the spatial resolution or MTF of the sensor. As the four-transistor structure is employed for each photodiode, eight transistors in total are required in a pixel, which slightly decreases the fill factor from 87% (calculated assuming a singlephotodiode pixel in the same dimension) to 85%. In reality, the maximum WDR performance is not only bounded by the noise level of each pixel but also the minimum exposure time. Long- and short-exposure data for a pixel are read from these two photodiodes at the same time and treated by their analog front-end circuit which consists of correlated double sampling (CDS) capacitors, sense amplifiers, and an 8-b ADC. Two photodiodes in a pixel are controlled separately by the timing generator. The integrations start at different times and end at the same time, thereby resulting in different integration times for both photodiodes. The start of the integration of a photodiode is done through its dedicated control signals, i.e., PTG1/2 and RST1/2. The readouts of dual-exposure data from both photodiodes are synchronized by the common signals such as row select (RSEL), column select (CSEL), signal data enable (SIG_IN), and reset data enable (RST_IN). To start the CDS process [9] at the end of integration, RST_IN is first asserted along with RSEL to store the reset data in the CDS capacitor C rst1/2, and then, SIG_IN and PTG1/2 are asserted to transfer the signal data to the CDS signal capacitor C sig1/2. When CSEL is asserted, the stored data are amplified by differential sense amplifiers and forwarded to dedicated ADCs. Because two different data of a pixel are available at the same time, no additional buffers are needed to synchronize them. On the other hand, a dual-exposure mechanism based on a single photodiode per pixel starts one exposure only after the other exposure is finished [6]. Accordingly, a considerable amount of digital memory is needed to store tens or hundreds of lines when an on-chip image synthesis is required. In addition, our approach is less vulnerable to distortion caused by camera or object motion than the scheme of a single photodiode per pixel. A block diagram, the physical layout, and a real sensor image of the implemented image sensor are shown in Fig. 11. In the block diagram shown in Fig. 11(a), gamma correction units are inserted between the imagers and the composer so that the user can enhance the final images by configuring their gammas by piecewise linear curves before images are merged in the composer unit. The Bayer format is utilized to achieve color images where one color filter is located on two photodiodes in a pixel. The Bayer color information on a pixel is maintained until color interpolation is performed to recover RGB data from each pixel in the Image Composer block. Other general ISP routines such as black level compensation, histogram stretching, postgamma correction, and color enhancement are applied to the composed output in the same block. Fig. 11(b) shows how the sensor components are sized and located in the sensor layout, which has 1/3-in optical area and an array of 7.8 μm 7.8 μm pixels and thus provides resolution. The vertical symmetry around the optical area shows that the sensor requires twofold analog circuits to treat long and short exposures separately. The sensor is delivered in a 48-pin CLCC package, as shown in Fig. 11(c). The maximum WDR capability of the sensor can be evaluated by its design parameters and pixel characteristics. Since the measured sensitivity of the long-exposure pixel is 1.6 V/lux s, the maximum illumination can be derived by the following: min_signal = 1.6 (5) min_lux max_exposure_time where min_signal is defined by the minimum signal level also represented in voltage and max_exposure_time is the maximum exposure time the sensor can provide to a pixel. Since the sensor produces 1 frame/s at the slowest mode and each pixel has 256 levels in 2.8-V operation voltage, by substituting

8 CHUN et al.: DYNAMIC-RANGE WIDENING IN CIS THROUGH EXPOSURE CONTROL OVER PIXEL 3007 help the sensor reproduce both the dark and the bright parts of a scene better as a single image. In this paper, it has been shown that a CIS integrated with its specialized exposure control over a dual-photodiode pixel and a simple synthesis technique can accomplish a 119-dB dynamic range without sacrificing the advantages of the conventional pixel structure. In spite of the increase of chip size due to the duplicated analog circuits, this approach provides improved performance compared with the previous dual-exposure-based works. Fig. 12. Images produced by the implemented sensor. (a) WDR function enabled. (b) WDR function disabled. min_signal and max_exposure_time with 2.8/256 V and 1 s, respectively, min_lux is given as lux. Similarly, since the sensitivity of the short-exposure pixel is measured as 5.4 V/lux s, the minimum illumination to determine the signal resolution will be given by the following: sat_level = 5.4 (6) max_lux min_exposure_time where sat_level refers to the peak voltage level of a pixel and min_exposure_time is the minimum exposure time for a pixel which is defined as the time for reading out one line at the maximum speed. By plugging values of 2.8 and (= (1/30(frames/s) 480 (lines))) into the parameters, respectively, max_lux will be 6341 lux. Therefore, the WDR capability of the sensor is calculated as 119 db (= 20 log(6341/0.0675)). A scene with a wide illumination distribution was photographed by the implemented sensor with the WDR function on and off, as shown in Fig. 12(a) and (b), respectively. When the WDR function is off, the sensor provides LEI only, and thus, the bright part is saturated in the result. It is shown that the saturated part is properly recovered in Fig. 12(b) by the dualexposure-based WDR function. VI. CONCLUSION A CIS based on a dual-photodiode pixel can widen the dynamic range of images if the exposure control of each photodiode is performed separately and adaptively according to its circumstance. Sizing the photodiodes differently can also REFERENCES [1] S. G. Chamberlain and J. P. Lee, Silicon imaging arrays with new photoelements, wide dynamic range and free from blooming, in Proc. Custom Integr. Circuits Conf., 1984, pp [2] T. Delbrück and C. A. Mead, Analog VLSI adaptive, logarithmic, widedynamic-range photoreceptor, in Proc. IEEE Int. Symp. Circuits Syst., 1994, pp [3] S. Decker, R. D. McGrath, K. Brehmer, and C. G. Sodini, A CMOS imaging array with wide dynamic range pixels and column-parallel digital output, IEEE J. Solid-State Circuits, vol.33,no.12,pp , Dec [4] D. X. D. Yang, A. El Gamal, B. Fowler, and H. Tian, A CMOS image sensor with ultrawide dynamic range floating-point pixellevel ADC, IEEE J. Solid-State Circuits, vol. 34, no. 12, pp , Dec [5] A. Bermak, A. Bouzerdoum, and K. Eshraghian, A vision sensor with on-pixel ADC and in-built light adaptation mechanism, Microelectron. J., vol. 33, no. 12, pp , Dec [6] O. Yadid-Pecht and E. R. Fossum, Wide intrascene dynamic range CMOS APS using dual sampling, IEEE Trans. Electron Devices, vol. 44, no. 10, pp , Oct [7] [Online]. Available: fujisuperccdsr.asp [8] B. C. Kuo, Automatic Control Systems. Englewood Cliffs, NJ: Prentice- Hall, 1995, pp [9] E. R. Fossum, CMOS image sensors: Electronic camera-on-chip, IEEE Trans. Electron Devices, vol. 44, no. 10, pp , Oct Jung-Bum Chun received the B.S degree in mathematics and the B.S degree in electronic engineering from Sogang University, Seoul, Korea, in 1996 and 1998, respectively, and the M.S. and Ph.D. degrees in electrical engineering from the Korea Advanced Institute of Science and Technology, Daejeon, Korea, in 2000 and 2009, respectively. From May 2000 to October 2004, he was with Paion Company, Ltd., Korea, where he developed high-speed Ethernet switching chipsets. Since November 2004, he has been developing CIS with Mtekvision Company, Ltd., and ClairPixel Company, Ltd., Seoul, Korea. Recently, he has been researching on image processing algorithms specialized for CIS and their optimal SoC implementation. Hunjoon Jung received the B.S. degree in physics and the Ph.D. degree in materials science and engineering from Seoul National University, Seoul, Korea, in 1986 and 2003, respectively. In January 1986, he joined LG Semiconductor Ltd. and worked on DRAM devices and CCD design for nine years. He is the holder of 26 patents on CCD. From 1995 to 1998, he was with Applied Materials Korea and worked on the development of semiconductor manufacturing equipment and processes. From 2003 to 2007, he was in charge of the Advanced Technology Center, Mtekvision Company, Ltd., Seoul, where he developed a CMOS image sensor, an LCD display driver, and a DRAM. Currently, he is the CEO of ClairPixel Company, Ltd., Seoul, which is developing novel CMOS image sensors and image signal processors.

9 3008 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 56, NO. 12, DECEMBER 2009 Chong-Min Kyung (S 76 M 81 SM 99 F 08) received the B.S. degree in electronics engineering from Seoul National University, Seoul, Korea, in 1975 and the M.S. and Ph.D. degrees in electrical engineering from Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea, in 1977 and 1981, respectively. From April 1981 to January 1983, he was with Bell Telephone Laboratories, Murray Hill, NJ, as a Postdoc. He is currently a Hynix Chair Professor with the Department of Electrical Engineering, KAIST, where he has been working on system-on-a-chip design, verification methodology, and processor and graphics architectures for high-speed and/or low-power applications, including the mobile video codec, since Dr. Kyung is a member of the National Academy of Engineering of Korea and the Korean Academy of Science and Technology. He was the General Chair of the 2007 IEEE Asian Solid-State Circuits Conference and the 2008 Asia and South Pacific Design Automation Conference (ASP-DAC). He was the recipient of the Most Excellent Design Award and the Special Feature Award in the University Design Contest in the ASP-DAC 1997 and 1998, respectively, and the Best Paper Awards in the 36th DAC, New Orleans, LA, the 10th International Conference on Signal Processing Application and Technology, Orlando, FL, in September 1999, and the 1999 International Conference on Computer Design, Austin, TX. He was also the recipient of a National Medal from the Korean government in 2000 for his contribution to research and education in IC design.