A 200X100 ARRAY OF ELECTRONICALLY CALIBRATABLE LOGARITHMIC CMOS PIXELS

Transcription

1 A 200X100 ARRAY OF ELECTRONICALLY CALIBRATABLE LOGARITHMIC CMOS PIXELS Bhaskar Choubey, Satoshi Aoyama, Dileepan Joseph, Stephen Otim and Steve Collins Department of Engineering Science, University of Oxford Oxford, OX1 3PJ, UK Renesas Technology Corp. Kodaira-shi, , Japan Department of Electrical and Computer Engineering, University of Alberta, Edmonton, T6G 2V4 Canada ABSTRACT CMOS logarithmic pixels are capable of capturing more than 6 decades of light intensity, whereas CCD and CMOS based Active Pixel Sensors saturate at about 3 decades. Logarithmic pixels also capture the contrast information of the image. However device variations result in fixed pattern noise in the response of these pixels. This noise can be modeled using a three parameters model. Offset based correction techniques, like double sampling and uniform stimulus based methods fail to produce a high quality image. In this paper an electronic calibration technique is used to calibrate a array of log pixels manufactured in AMS 0.35 micron technology. Each column of pixels has a current source, which is used to feed two different current through each pixel. The response of each pixel to two input currents together with the dark response are used to extract the three parameters for every pixel. These are used to correct the final image. The residual fixed pattern noise after the correction is found to match performance of the human eye. 1. INTRODUCTION The human visual system has the ability to interpret scenes with a wide range of varying illuminations from lux to those as bright as lux at a minimum 1% contrast sensitivity in high illumination [1]. Typical real world scenes have dynamic intra-scene ranges that might extend about five orders of magnitude, from 1 lux in shadows to lux of bright sunlight. Unfortunately, Charged Coupled Devices (CCD s and CMOS Active Pixel Sensors (APS, which currently dominate the image sensor market, have a dynamic range of less than three orders of magnitude. Consequently, when imaging a wide dynamic natural scene the response of these sensors saturates in some regions of the scene. Several techniques such as multi-sampling, stepped reset voltages, threshold comparison have been proposed that can extend the dynamic range of these sensors [2], [3], [4]. However most of these techniques have complex inpixel circuit, leading to low fill factor. In addition, any linear sensors would require large number of bits, to encompass the wide dynamic range of scenes. Logarithmic image sensors designed in CMOS technology using subthreshold region of operation of a MOS are capable of capturing wide dynamic range scenes, with intensity variations of more than 6 decades [5, 6]. These sensors provide similar fill factor to that of CMOS Active Pixel Sensors, as well as random addressability. Another potential advantage of logarithmic pixels is that they encode the contrast information from a scene. Figure 1 shows a conventional logarithmic pixel. It consists of a photodiode to convert input illumination to a photocurrent. This current is converted into a voltage in a logarithmic fashion by the subthreshold device M1. M2 is a source follower, while M3 is a row select switch. There are other readout circuits in the column generally involving current source for the source follower and another source follower to select between various columns. PD Vdd M1 N1 RS M2 M3 Fig. 1. A logarithmic Pixel Readout Circuits

2 Variations between devices within different pixels causes this type of sensor to suffer from fixed pattern noise(fpn, which severely degrades the quality of the resulting image. Various techniques have been used to reduce the effect of fixed pattern noise on the image. Ricquer et.al. have used hot electron injection within each pixel to alter the threshold voltage of the load transistor [7] to reduce offset variations. However, even when high stressing voltages are applied to increase the number of hot electrons available this proved to be a time consuming process. In an alternative approach Loose and coworkers [8], have used feedback to adjust the gate voltage of each load transistor. However the additional circuitry required to use this technique increases the pixel area and reduces fill-factor. The reported pixel has a fill factor of 30% in a pixel of dimension chee24µ 24µm. Kavadias et. al have proposed another technique in which additive FPN is reduced by subtracting a known response of the pixel from its image response. This reference response is obtained by feeding particular a high current to the pixel [5]. But the residual FPN was very high to produce a good quality image. Various off-chip methods have been also been proposed which involve storing sensor s response to uniform scene is stored and used to remove additive FPN by subtracting from an image response [9, 10] We have shown previously that offset only calibration is insufficient to meet requirement of a high quality image. We proposed a 2 parameter calibration scheme to reduce the fixed pattern noise caused by offset and gain. The scheme was proved experimentally in a column of 200 pixels [6]. In this paper, we are extending the principle to an array pixel. The rest of paper is arranged as follow. Section 2 described the model and related FPN correction technique. The experimental chips and setups are explained in section 3. Section 4 discusses the experimental setups and results. Section 5 provides the conclusion. 2. ELECTRONIC CALIBRATION By considering the characteristics of the devices within a logarithmic pixel, Joseph and Collins have shown that the response of a logarithmic pixel, y to a photocurrent x, can be written in the form[11] y = a + bln(c + x (1 Here a is an additive offset, b is the gain of the photodetector, and c is a leakage current based bias. Offset a depends on the threshold voltage of various transistor and hence forms the dominant source of fixed pattern noise. The gain parameter b, depends on the gain of the readout circuits and body effect and hence contribute to fixed pattern noise. Bias parameter c, that represents the effects of a leakage current within the pixel, mainly from the photodiode. An important effect of this third term is to limit the sensitivity of each pixel at low illumination levels. At higher photocurrent the effects of this term are negligible. In order to correct for variations in the offset, gain and bias of each pixel a technique is required to determine the value of these three parameters for each pixel. When demonstrating the validity of the three parameter model Joseph and Collins used data from twenty four uniform images at different illumination levels and an iterative parameter extraction technique [11]. It is impractical to use this technique to obtain the data required to correct an image for fixed pattern noise, as it is computationally demanding as well as requires 24 images at different illumination intensity. However, there are only three parameters in the model and these parameters can therefore be estimated using three data points per pixel. Obtaining these data points electronically will make fixed pattern noise correction convenient for the user. The other advantage of electronic calibration is that the reference currents used to obtain the data for parameter extraction can be selected to simplify the parameter extraction procedure. For example, for the calculation of the offset and gain of each pixel, two of the reference currents should be chosen so that they are much larger than the leakage current in each pixel. Under these conditions the contribution of c becomes negligible. Then if the pixel response at reference current x 1 is y 1 and its response to another current is y 2 the two parameters can be calculated using b = y 1 y 2 ln(x 1 / (2 a = y 1 bln(x 1 (3 The value of the third parameter c is then best determined using a data point from the operating region in which the leakage current is larger than the photocurrent. Therefore the best data to use to obtain this parameter is when the only current flowing through the load transistor is the leakage current. This data point corresponds to the dark response, y d of the pixel and since at this point x = 0 c = exp( y d a (4 b This parameter can then be used to characterise the minimum illumination level at which the pixel gives a logarithmic response. Using the three parameters, value of photointensity at every pixel (x ij for pixel of ith row and jth column site can be computed, using x ij = exp( y ij a ij b ij c ij (5 The accuracy of the three parameter model when equations 2, 3 and 4 have been used to extract the parameters has been investigated by Otim and coworkers using data obtained from a circuit simulator [12]. They simulated a logarithmic pixel manufactured on a standard 0.35 micron CMOS process

3 with device geometries chosen to represent a 10µm by 10µm pixel. They have found that relative contrast error of the extracted image of an uniform scene assuming ideal readouts, have been as low as 0.5%. In this calibration process, computation of three parameters requires knowledge of exact values of calibration currents, it can be shown that the residual fixed pattern noise is independent of these currents and depends only on ratio of the calibration currents. Expressing the various parameters in equation 5 in terms of pixel s calibration currents and calibration voltage, ( yij (y ij1 b ij log(x 1 x ij = exp c ij b ij ( yij y ij1 = x 1 exp + x 1 exp log( x 1 y ij1 y ij2 ( yijd y ij1 log( x 1 y ij1 y ij2 (6 = x 1 f(y ij,y ij1,y ij2,y ijd, x 1 (7 The residual fixed pattern noise of such an array expressed as ratio of standard deviation of extracted currents to average of extracted currents, will then be independent of x 1, depending only on ratio of two calibration currents, as follow. std(x ij mean(x ij = std(f(y ij,y ij1,y ij2,y ijd, x1 mean(f(y ij,y ij1,y ij2,y ijd, x1 3. CHIP DESIGN The logarithmic pixel that has been designed to prove the calibration process is shown in figure 2. In this circuit transistor M1 is the load device that converts the photocurrent to a voltage. Transistor M4 was included in the pixel to limit the voltage across the photodiode while transistor M5 acts as a switch. This switch can be used during calibration to selectively connect the pixel to the drain of transistor M6 which acts as the voltage controlled current source. In order to save area and to ensure uniformity of the current flowing in different pixels, this device is shared by all the pixels in the same column. In a conventional logarithmic pixel a source-follower circuit is used as the readout circuit. The problem with this simple circuit is that it has a relatively low gain. This can be a problem for logarithmic pixels which typically have a sensitivity of approximately 70 mv/decade. In order to minimise any attenuation of the resulting small signals by the readout circuit the source-follower has been replaced by a differential amplifier. In this circuit, the two pmos transistors form a current mirror and transistor M8 acts as a constant current source. Assuming that the on resistance of (8 transistor M3 is zero and that the current mirror is ideal, then the gate-source voltages of transistors M2 and M2 will be equal. In these circumstances this circuit acts as a voltage follower. This circuit therefore can have a significantly higher gain than the alternative source-follower circuit. In addition, since only transistors M2 and M3 are actually within each pixel, this increased gain can be achieved without increasing the area or variability of the pixels. Calssel Vcs M6 Vcasc M5 Pixel Vdd M1 M4 Row Select PD M7 Column Bias Vdd M2 M2 M3 M8 M7 Column Voltage Fig. 2. A Logarithmic Pixel with Calibration Current Source and Differential Amplifier Readout The chip was manufactured in Austrian Mirosystems 0.35 µ, 2-poly, 3-metal process. It had a 200 x 100 array pixels of dimension 10 µ 10µ. N-diff-P-substrate photodiode was used. In addition to pixels, the array also had column and row selection circuitry, which involved a 200 and a 100 bit shift register. The Setup 4. EXPERIMENTAL PROCEDURE All clocks were generated externally by instrunet data acquisition board which in turn was controlled by Agilent Vee programming interface. The pixel values were read in an analogue fashion, through a semiconductor parameter analyzer. The program used to generate the clock signals was also used to collect data from semiconductor parameter analyzer using GPIB connection. Many readings were taken at every point to reduce the effect of temporal noise. To remove transient effects, a very slow speed clock as low as 1 Hz was used for operation. Further all electronic experiments were carried out in a stable temperature oven to minimise temperature variations. The current source in figure 2

4 was used to determine the response of each pixel when its load transistor was forced to source two different current. These two measured responses from each pixel were then used to determine both the offset and the gain of each pixel. The dark point was used to compute the bias parameter. FPN Correction With this information the responses of the pixels at several other currents, over a range of more than six decades, have been corrected in order to reduce the fixed pattern noise. Figure 3 shows the residual fixed pattern noise at different currents fed by the current source. It is observed that fixed pattern noise is reduced to less than 2% contrast threshold for over 6 decades of currents corresponding to six decades of light intensity. Figure 4 shows the residual fixed pattern noise when only 1 parameter was corrected by subtracting the pixel response from its reference response, in a fashion similar to [5, 9], showing higher residual error is seen in these frames. The uncorrected fixed pattern noise was found to be of the order of 104%. The residual contrast in uncorrected and offset corrected image is detoriated to such high levels, as the distribution of extracted intensity(represented by photocurrents is not gaussian, rather is log-normal. Contrast Sensitivity Contrast Sensitivity Photocurrents in A Fig. 4. Contrast Sensitivity expressed as Standard Deviation as a percentage of mean of extracted currents with offset correction correction hand the second point should not be very far from the first point, as it would bring it might bring it near the moderate inversion region, wherein the model may fail. Otim and coworkers have made a detailed study of these effects on the modeling of logarithmic pixels [12]. They show that calibration points should be at least two decades distant from the dark point and have about 2-3 decades separation between them, for optimum minimal FPN. Figure 5 compares the residual fixed pattern noise, expressed as contrast sensitivity, when the array of pixels was excited electronically at constant temperature. One of the calibration points has been fixed while the second has been made variable. It is seen that the residual error in between the calibration region as well as those outside it, depends very much on the position of calibration points Photocurrents in A 10 9 Fig. 3. Contrast Sensitivity expressed as Standard Deviation as a percentage of mean of extracted currents with 2 parameter correction Calibration Points An important consideration with such a calibration technique is proper selection of calibration points. The two calibration points needed to compute offset and gain value of each pixel, should be considerably apart from the dark point so as to minimise the role of leakage current. The two calibration points should also have considerable distance apart as well. They should not be near enough for temporal noise to considerably influence the gain parameter. On the other Contrast Sensitivity Photocurrent in A Fig. 5. Effect of change of calibration point to residual fixed pattern noise

5 Optical Stimulation To compare the performance with optical simulation at room temperature, the chip was exposed to a uniform scene. The uniform scene was generated by using a white paper. The camera lens was focused at infinity and thus a diffused surface was obtained. The array was excited using optical as well as electronic current of similar values in successive experiments. It was observed that residual fixed pattern noise is higher in case of optical data then that of electronic data. This suggests that photodiodes are having variations in their response to uniform scenes. This form of variation brings in another source of fixed pattern noise, which is evident from the values of residual fixed pattern noise. Despite increase in the residual FPN, contrast sensitivity values are still within 2 to 3%, the response limits of the human vision. The 200x100 image sensor was used with the calibration process to correct various uniform scenes at various light intensities. Figures 6, 7 and 8 shows the image of an uniform scene produced by the sensor without any correction, without single parameter correction and with three parameter correction. Figure 9 shows the residual image formed with electrical simulation of similar nature as that of optical simulation 1. It was found that a worst case of 3% contrast threshold was obtained for about 6 decades of illumination. This corresponds to about 1.28% residual fixed pattern noise expressed as a percentage of one decade of illumination. In the best case scenario, near the calibration points, the residual fixed pattern noise was found to be as low as 1.5% contrast threshold. Fig. 7. Optical Uniform Scene with offset correction only Fig. 8. An optical Uniform scene with 3 parameter correction Fig. 9. An Electrically generated image with 3 parameter correction Fig. 6. Uncorrected Optical Uniform Scene 5. CONCLUSION Logarithmic pixels have the ability to match the performance of human eye as far as the high dynamic range is concerned, but they are crippled by fixed pattern noise, which appear due to device to device variations. Several techniques have 1 The authors are aware that the print of the paper may not be able to show the difference in images due to printing limitations. These images in raw format can be obtained from the authors. Table 1. Chip Statistics Technology Austrian Microsystems 0.35µ Array Pixel 5-Transistor Size 10µ 10µ Fill Factor 49% Readout 2 stage differential Dynamic Range 9decades Average FPN 3% Contrast threshold

6 been used in literature to rectify the image from FPN, most of which concentrate only on additive FPN, resulting in high residual noise. In this paper an electronic calibration technique has been shown to be an efficient way to correcting most of the fixed pattern noise. Residual error as low as 2% has been obtained in the imager s response to uniform scenes, which matches the performance of human eye. 6. REFERENCES [11] D.Joseph and S. Collins. Modelling, calibration and correction of illumination-dependent fixed pattern noise in logarithmic cmos image sensor. In IEEE Trans. Inst. and Meas., 51(5, October [12] S. Otim, D. Joseph, B. Choubey and S. Collins. Model Based Fixed Pattern Noise Correction for High Dynamic Range Logarithmic CMOS Imagers. submitted for publication [1] B.A. Wandell Foundations of Vision Sinauer Associates, 1995 [2] O. Yadid-Pecht and E. R Fossum., Image sensor with ultra high linear dynamic range utilising dual output cmos active pixel sensors, IEEE Transactions on Electron Devices, vol. 44, no. 10, pp , October [3] A. Bubmann B J Hostika M. Schanz, C. Nitta and R K Wertheimer., A high dynamic range cmos image sensor for automotive applications., IEEE Journal on Solid State Circuits, vol. 35, no. 7, pp , July [4] L. Gonzo D. Stopa, A. Simoni and G. Dalla Betta., Novel cmos image sensor with a 132-db dynamic range., IEEE Journal on Solid State circuits, vol. 37, no. 12, pp , December 2002 [5] D. Scheffer S. Kavadias, B. Dierickx, A. Alaerts, D. Uwaerts, and J. Bogaerts. A Logarithmic Response CMOS Image Sensor with On-Chip Calibration. IEEE JSSC, 35(8: , August [6] B. Choubey, S. Ayoma, D. Joseph, S. Otim and S. Collins. A 3 Parameter Electronic Calibration Scheme For Logarithmic CMOS Pixels. In IEEE IS- CAS, May [7] N.Ricquer and B. Dierickx. Active pixel cmos sensor with on chip nonuniformity correction. In Proc. IEEE Workshop on CCD and AIS, [8] M. Loose, K. Meier, and J. Schemmel. A Self- Calibrating Single-Chip CMOS Camera with Logarithmic Response. IEEE JSSC, 36(4:586 96, April [9] IMS Chips. HDRC VGA Imager and Camera Data and Features. Technical report, Institute for Microelectronics Stuttgart, September [10] G. F. Marshall and S. Collins. A High Dynamic Range Front End for Automatic Image Processing Applications. In Proceedings of the SPIE, volume 3410, pages , May Advanced Focal Plane Arrays and Electronic Cameras II.