High Dynamic Range Image Sensors

Abbas El Gamal

Department of Electrical Engineering Stanford University

ISSCC02

Motivation

Some scenes contain very wide range of illumination with intensities varying over 100dB range or more Biological vision systems and silver halide lm can image such high dynamic range scenes with little loss of contrast information Dynamic range of solid-state image sensors varies over wide range: high end CCDs > 78dB consumer grade CCDs 66dB consumer grade CMOS imagers 54dB So, except for high end CCDs, image sensor dynamic range is not high enough to capture high dynamic range scenes

ISSCC02

Imaging High Dynamic Range Scene The problem

HDR Scene

Short Exposure-time Image

Medium Exposure-time Image

Long Exposure-time Image

ISSCC02

Extending Sensor Dynamic Range

Several techniques and architectures have been proposed for extending image sensor dynamic range many questions: What exactly is sensor dynamic range, what does it depend on, and how should it be quantied (100dB, 12 bits, . . . )? In what sense do these techniques extend dynamic range? How well do they work (e.g., accuracy of capturing scene information and/ or image quality)? How should their performance be compared? Are all 100dB sensors equivalent? Is a 120dB sensor better than a 100dB one?

ISSCC02

Tutorial Objectives

To provide quantitative understanding of sensor DR and SNR and their dependencies on key sensor parameters To describe some of the popular techniques for extending image sensor dynamic range To provide a framework for comparing the performance of these techniques: Quantitatively based on DR and SNR Qualitatively based on other criteria, e.g., linearity, complexity . . .

ISSCC02

Outline

Background
Introduction to Image Sensors Image Sensor Model Sensor Dynamic Range (DR) and SNR

Description/ Analysis of HDR Schemes Other HDR Image Sensors Conclusion References

ISSCC02

Image Sensors

Area image sensor consists of: 2-D array of pixels, each containing a photodetector that converts light into photocurrent and readout devices Circuits at periphery for readout, . . . Since photocurrent is very small (10s 100s of fA), it is dicult to read out directly Conventional sensors (CCDs, CMOS APS) operate in direct integration photocurrent is integrated during exposure into charge, read out as voltage In some high dynamic range sensors photocurrent is directly converted to output voltage

ISSCC02

Signal Path for Single Pixel

Photons
ph/s

Photocurrent
Amp

Charge
Col

Voltage
V

Photon ux conversion to photocurrent is linear (for xed irradiance spectral distribution) governed by quantum eciency Photocurrent is integrated into charge during exposure In most sensors, charge to voltage conversion is performed using linear amplier(s)

ISSCC02

Direct Integration
vReset Q reset vo i CD tint t high light Qsat low light

Direct integration: The photodetector is reset to vReset Photocurrent discharges CD during integration time or exposure time, tint At the end of integration, the accumulated (negative) charge Q(tint) (or voltage vo(tint)) is read out Saturation charge Qsat is called well capacity

ISSCC02

Interline Transfer CCD Image Sensor

Vertical CCD

Integration Transfer Vertical shift

Photodiode

Horizontal shift Time Operation

Horizontal CCD Output Amplifier

Collected charge is simultaneously transferred to the vertical CCDs at the end of integration time (a new integration period can begin right after the transfer) and then shifted out Charge transfer to vertical CCDs simultaneously resets the photodiodes shuttering done electronically (see [2])
ISSCC02 10

CMOS Active Pixel Sensor (APS)

Pixel (i,j) Reset i

Integration Time
Word i Reset i Col j

Word i
Row i Read

Bit j

Word i+1 Reset i+1

Col j Out

Vbias

Coj Column Amp/Mux

Operation

Pixel voltage is read out one row at a time to column storage capacitors, then read out using the column decoder and multiplexer Row integration times are staggered by row/column readout time

ISSCC02

11

Image Sensor Non-idealities

Temporal noise Fixed pattern noise (FPN) Dark current Spatial sampling and low pass ltering

ISSCC02

12

Temporal Noise

Caused by photodetector and MOS transistor thermal, shot, and 1/f noise Can be lumped into three additive components: Integration noise (due to photodetector shot noise) Reset noise Readout noise Noise increases with signal, but so does the signal-to-noise ratio (SNR) Noise under dark conditions presents a fundamental limit on sensor dynamic range (DR)

ISSCC02

13

Fixed Pattern Noise (FPN)

FPN is the spatial variation in pixel outputs under uniform illumination due to device and interconnect mismatches over the sensor Two FPN components: oset and gain Most visible at low illumination (oset FPN more important than gain FPN) Worse for CMOS image sensors than for CCDs due to multiple levels of amplication
FPN due to column amplier mismatches major problem

Oset FPN can be reduced using correlated double sampling (CDS)

ISSCC02

14

Dark current

Dark current is the leakage current at the integration node, i.e., current not induced by photogeneration, due to junction (and transistor) leakages It limits the image sensor dynamic range by introducing dark integration noise (due to shot noise) varying widely across the image sensor array causing xed pattern noise (FPN) that cannot be easily removed reducing signal swing

ISSCC02

15

Sampling and Low Pass Filtering

The image sensor is a spatial (as well as temporal) sampling device frequency components above the Nyquist frequency cause aliasing It is not a point sampling device signal low pass ltered before sampling by spatial integration (of current density over photodetector area) crosstalk between pixels Resolution below the Nyquist frequency measured by Modulation Transfer Function (MTF) Imaging optics also limit spatial resolution (due to diraction)

ISSCC02

16

Image Sensor Model

Photocurrent to output charge model: QShot idc iph i

QReset QReadout QFPN

Q(i)

Qo

Q(i) is the sensor transfer function and is given by: Q(i) =

1 q (itint )

Qsat

electrons for 0 < i < qQsat tint qQsat for i tint

QShot is the noise charge due to integration (shot noise) and has average power 1 (iph + idc)tint electrons2 q QReset is the reset noise (KT C noise)
QReadout is the readout circuit noise QFPN is the oset FPN (we ignore gain FPN) All noise components are independent
ISSCC02 17

Input Referred Noise Power

To calculate SNR and dynamic range we use the model with equivalent input referred noise current idc In

iph

Q(.)

Qo

Since Q(.) is linear we can readily nd the average power of the equivalent input referred noise r.v. In, i.e., average input referred noise power, to be q2 1 2 2 In = 2 ( (iph + idc)tint + r ) A2, tint q where 2 2 2 2 r = Reset + Readout + FPN electron2, is the read noise power
ISSCC02 18

Correlated Double Sampling (CDS)

The output of the sensor is sampled twice; once right after reset and the second time with the signal present Sample without signal: QReset = QReset + QReadout + QFPN o Sample with signal: 1 Qo = (iph + idc)tint + QShot + QReset + QReadout + QFPN q The dierence is: 1 Qo QReset = (iph + idc)tint + QShot + (Q Q )Reset + (Q Q )Readout o q
oset FPN is eliminated, readout noise power is doubled, and reset noise is either eliminated (CCDs, photogate APS) or doubled (photodiode APS) WARNING: If sensor transfer function Q(.) is not linear, CDS DOES NOT WORK
ISSCC02 19

Signal to Noise Ratio (SNR)

SNR is the ratio of the input signal power to the average input referred noise power, and is typically measured in dBs Using the average input referred noise power expression, we get SNR(iph) = 10 log10 i2 ph
q2 1 ( (i t2 q ph int

+ idc)tint +

2 r )

dB

SNR increases with the input signal iph, rst (for small iph) at 20dB per decade since read noise dominates, then at 10dB per decade when shot noise (due to photodetector) dominates

ISSCC02

20

SNR Versus iph (for three idc values)

50

Qsat = 40000 e r = 20 e
40

tint =10 ms

30

SNR (dB)

20

10

0
1fA 5fA 15fA

10 16 10

10

15

10

14

10

13

10

12

10

11

iph (A)
ISSCC02 21

Dynamic Range
Dynamic range quanties the ability of a sensor to adequately image both high lights and dark shadows in a scene It is dened as the ratio of the largest nonsaturating input signal to the smallest detectable input signal largest nonsaturating signal given by imax = is the well capacity
qQsat tint

idc, where Qsat

smallest detectable input signal dened as standard deviation of input referred noise under dark conditions In (0) (the zero here q 2 refers to iph = 0), which gives imin = tint 1 idctint + r q

Thus dynamic range imax DR = 20 log10 = 20 log10 imin

qQsat tint q tint

idc
2 + r

1 q idc tint

ISSCC02

22

Dynamic Range Versus Integration Time

70

68

66

64

62

DR (dB)

60

58

56

54

Qsat = 40000 e r = 20 e idc = 1 fA

10
3

52

50 4 10

10

10

Integration time tint (sec)

ISSCC02 23

SNR vs. iph for Dierent Integration Times

50

Qsat = 40000 e r = 20 e
40

40ms

20ms

10ms

5ms

idc = 1 fA

30

SNR (dB)

20

10

10 16 10

10

15

10

14

10

13

10

12

10

11

iph (A)
ISSCC02 24

Summary

Presented brief tutorial on image sensor operation and nonidealities Described sensor signal and noise model Used the model to quantify sensor SNR and DR Discussed dependencies of SNR and DR on sensor parameters

ISSCC02

25

Outline Background Description/ Analysis of High DR Schemes

Well Capacity Adjusting Multiple Capture Spatially Varying Exposure Time to Saturation

Other HDR Image Sensors

Logarithmic Sensor Local Adaptation Conclusion

References
ISSCC02 26

Extending Dynamic Range

To increase dynamic range we need to increase imax and/or decrease imin
imax imin =
qQsat tint ,

increases as integration time is decreased + ( qr )2, decreases as integration time is increased tint

q tint idc

To increase dynamic range need to spatially adapt pixel integration times to illumination
short integration times for pixels with high illumination long integration times for pixels with low illumination

Integration time cant be made too long due to saturation and motion The HDR techniques only increase imax (for a given maximum integration time)
Recent work [3] shows how imin can be reduced by lowering read noise and preventing motion blur

ISSCC02

27

The Plan

To describe Well Capacity Adjusting and Multiple Capture To analyze their SNR and show that the increase in dynamic range comes at the expense of decrease in SNR
Multiple Capture achieves higher SNR than Well Capacity Adjusting for the same increase in dynamic range (see [4])

To describe two other techniques: Spatially Varying Exposure and Time-to-Saturation To briey describe two other types of sensors that do not use direct integration: Logarithmic Sensor and Local Adaptation To qualitatively compare these six HDR techniques

ISSCC02

28

Well Capacity Adjusting

Available well capacity is increased one or more times during integration (idea initially applied to CCDs [5,6]) For APS this is done by adjusting the reset signal one or more times during integration [7]: vReset vdd
Reset Follower

vdd

tint
Access

Bias

ISSCC02

29

Sensor Transfer Function

The current to charge transfer function is compressed resulting in higher maximum nonsaturating current imax: Q(i) Qsat

imax

imax

ISSCC02

30

Analysis of Well Capacity Adjusting

Consider the case of a single well capacity adjustment: Q Qsat
highest light

moderate light

Qsat
low light

t1

tint

Largest nonsaturating current is now given by (1 )qQsat imax = idc tint t1 Smallest detectable signal does not change, so dynamic range is increased by a factor (1 ) DRF 1 (1 ttint )
ISSCC02 31

Sensor Transfer Function

Q(i) Qsat

Qsat tint t1

qQsat t1

qQsat (1) (tint t1 )

ISSCC02

32

Well Capacity Adjusting Example

78dB Scene

Using Well Capacity Adjusting (DRF=16)

ISSCC02

33

SNR for Well Capacity Adjusting

Again consider the single well capacity adjusting case: For 0 iph < qQtsat idc, the SNR is the same as for the normal sensor 1 operation, and we get
SNR(iph) = 10 log10

i2 ph
q2 1 ( (i t2 q ph int

+ idc)tint +

2 r )

dB

(1) For qQtsat idc iph < qQsatt1) idc, the variance of the noise charge is (tint 1 1 2 given by q (iph + idc)(tint t1) + r electrons2

So the variance of the equivalent input referred noise current is given by 2 2 2 In = (t q )2 ( 1 (iph + idc)(tint t1) + r ) A2 q t
int 1

Thus, the signal to noise ratio SNR(iph) = 10 log10 i2 ph

q2 1 2 ( q (iph (tint t1 )

+ idc)(tint t1) +

2 r )

dB

ISSCC02

34

Example
Qsat = 40000 e
45

r = 20 e tint = 10 ms idc = 1 fA = 3 , t1 = 9.8438 ms 4 DRF = 16

40

35

30

SNR (dB)

25

20

15

10

10

15

10

14

10

13

10

12

10

11

iph (A)
ISSCC02 35

SNR DIP
Notice the 21dB dip in the SNR example at the transition point: iph = qQsat idc = 487 fA t1

In general, the DIP is given by

2 (Qsat (tinttt1) + r ) t2 int 1 DIP = 10 log10 2 (tint t1)2 (Qsat ttint + r ) 1 1 10 log10 1 1 ttint DRF = 10 log10 (1 )

Thus it increases with DRF, but also as increases

ISSCC02

36

Well-Adjusting Issues

Increasing DR directly lowers SNR Implementation is straightforward well adjusting causes more noise and FPN (not accounted for in analysis) Sensor response is nonlinear CDS only eective at low illumination Color processing ?

ISSCC02

37

Multiple Capture
Idea: Capture several images within normal exposure time
short integration time images capture high light regions long integration time images capture low light regions

Combine images into HDR image, e.g., using Last Sample Before Saturation algorithm:
Only extends dynamic range at high illumination end

Implementation of 2 captures demonstrated for CCDs and CMOS APS [8] Implementing many captures requires very high speed non-destructive readout 9 captures demonstrated using DPS [9] Recent work [3] shows that dynamic range can also be extended at the low illumination by appropriately averaging the captured images to reduce read noise
ISSCC02 38

Multiple Capture Example

T 2T 4T

8T
ISSCC02

16T

32T
39

High Dynamic Range Image

ISSCC02

40

Analysis of Multiple Capture

Consider the case of two captures: Q Qsat
highest light

moderate light

low light tint a

tint

The largest nonsaturating current is now aqQsat idc imax = tint The smallest detectable signal does not change, so dynamic range is increased by a factor DRF a
ISSCC02 41

Sensor Transfer Function

The current to charge response using the Last Sample Before Saturation algorithm Q Qsat

Qsat a

qQsat tint

aqQsat tint

ISSCC02

42

SNR for Multiple Capture

Consider the case of two captures and the Last Sample Before Saturation algorithm For 0 iph < qQsat idc, the SNR is the same as for the normal tint operation, and we get
SNR(iph) = 10 log10

i2 ph
q2 1 ( (i t2 q ph int 2 + idc)tint + r )

dB

For qQsat idc iph < aqQsat idc, the SNR is the same as normal tint tint operation with tint replaced by tint/a, and we get SNR(iph) = 10 log10 i2 ph
a2 q 2 1 ( q (iph t2 int

idc) tint a

2 r )

dB

ISSCC02

43

SNR Example
50

Qsat = 40000 e r = 20 e
40

tint = 10 ms idc = 1 fA DRF = 16

30

SNR (dB)

20

10

Well Multi

10 16 10

10

15

10

14

10

13

10

12

10

11

10

10

iph (A)
ISSCC02 44

SNR DIP
Notice that SNR dips by 12.6 dB at the transition current iph = qQsat idc = 640 fA tint In general
2 sat a2( Qa + r ) 10 log10 DRF dB DIP = 10 log10 2 (Qsat + r )

Recall that for well capacity adjusting DIP 10 log10 DRF , (1 )

which is larger for the same DRF

ISSCC02

45

Well Adjusting versus Multiple Capture

50

Qsat = 40000 e r = 20 e
40

tint =10ms idc = 1fA DRF=256

30

SNR (dB)

20

10

8 adjustments for well adjusting (solid) 9 captures for multiple capture (dashed)
10
15

10 16 10

10

14

10

13

10

12

10

11

10

10

10

iph (A)
ISSCC02 46

Multiple Capture Issues

SNR for normal sensor signal range is not aected SNR maintained over extended dynamic range Response is linear can use CDS and perform conventional color processing Dual capture implementation not too dicult (plenty of time between the two captures to readout the rst capture) High SNR dip for regions of moderate light HDR image can look segmented Multiple capture implementation is quite dicult, needs Very high speed non-destructive readout (see [10]) On-chip memory and some logic to perform reconstruction of HDR image during capture

ISSCC02

47

Spatially Varying Exposure

The spatially varying exposure technique [11] implements multiple capture using a conventional sensor by sacricing spatial resolution The idea is to deposit an array of neutral density lters on the sensor so that in a single capture pixels with darker lters sample high lights while pixels with lighter lters sample low lights

High dynamic range image synthesized using low pass ltering or more sophisticated techniques such as cubic interpolation

ISSCC02

48

Spatially Varying Exposure Issues

Very simple to implement and requires no change to the sensor itself Blocking light reduces sensor sensitivity and SNR Very high resolution sensor is needed, since spatial resolution is reduced DR is extended at the high illumination end only (same as multiple capture using the Last Sample Before Saturation algorithm)

ISSCC02

49

Time-to-Saturation
The idea is to measure the integration time required to saturate each pixel, the photocurrent is estimated by qQsat tsat so the sensor current to time transfer function is nonlinear iph =
SNR is maximized for each pixel and is equal to SNR = 20 log10

Qsat 10 log10 Qsat 2 Qsat + r

Minimum detectable current is limited by maximum allowable integration time (not zero !) qQsat imin tint and the maximum detectable current is limited by circuit mismatches, readout speed, and FPN (it is not innity !) [12,13,14]

ISSCC02

50

Time-to-Saturation Issues

Implementation is quite dicult need a way to detect saturation for each pixel, then record the time If global circuits are used [13], contention can severely limit performance If done at the pixel level [14], pixel size may become unacceptably large imin can be unacceptably large (=213fA for Qsat = 40000e and tint = 30ms) unless sensor is read out at tint No CDS support

ISSCC02

51

Logarithmic Sensor
In logarithmic sensors, photocurrent is directly converted to voltage for readout [15] vdd

iph

vout

High dynamic range achieved via logarithmic compression during conversion (to voltage), using exponential IV curve of MOS transistor in subthreshold iph vout = k ln Io Up to 56 decades of dynamic range compressed into 0.5V range (depending on vT and the number of series transistors)
ISSCC02 52

MOSFET Subthreshold I-V Characteristic

10 10 10 10

ids (A)

10 10 10 10 10 10

10

11

12

13

Subthreshold Region

14

15

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

vgs (V)
ISSCC02 53

Logarithmic Compression Issues

Large variation in subthreshold MOSFET characteristics and threshold voltage cause high FPN
Oset FPN as high as 10-40% of swing reported Since CDS cannot be used, FPN can only be corrected by storing the osets or other techniques, which increase complexity [15]

Succeeding circuitry must be extremely precise to make use of the dynamic range aorded by the compressed output voltage Non-integrating nature limits achievable SNR even for high illumination due to exponential transconductance relationship i2 kT CD kT CD ph 10 log10 SNR(iph) = 10 log10 2 q (iph + idc)2 q2
Note: Here we neglected read noise (which would make SNR even worse)

ISSCC02

54

SNR for Logarithmic vs. Integrating Sensor

45

Qsat = 40000 e r = 20 e tint = 20ms idc = 1fA

40

35

30

SNR (dB)

25

20

15

10

Integrating Logarithmic

10

15

10

14

10

13

10

12

10

11

10

10

10

iph (A)

ISSCC02

55

Local Adapatation (Articial Retina)

The idea is to mimic biological vision systems where local averages are subtracted o each receptor value in order to increase the contrast while maintaining wide dynamic range [16,17] Hexagonal resistive network is used to weight closer pixels exponentially higher than distant ones, thereby approximating Gaussian smoothing

vout

ISSCC02

56

Comparison Between HDR Schemes

Well-Adjust Multi-Capture Spatial Saturation Log Adapt DR: imax imin Linearity CDS SNR Complexity: Pixel External Same No No Low Low Same/ Yes Yes + Med-High Low-Med Yes Yes Low High No No High High No No No No

Med High Low Low

ISSCC02

57

Conclusion
Dened image sensor DR and SNR and discussed their dependencies on sensor parameters Described six techniques for extending DR:
Well Adjusting Mutiple Capture Spatially Varying Exposure Time to saturation Logarithmic Local adaptation (articial retina)

All techniques only increase imax (Multiple Capture can also decrease imin by weighted averaging)
SNR can be used to compare dierent HDR sensors Multiple Capture with weighted averaging is the most eective method but requires high speed non-destructive readout sensor and DSP hardware to implement
ISSCC02 58

References
[1] A. El Gamal, 392b Lectures and Classnotes, Stanford University, 2001. [2] A. J. P. Theuwissen, Solid-State Imaging with Charge-Coupled Devices. Norwell, MA: Kluwer, 1995. [3] X.Q. Liu and A. El Gamal, Simultaneous Image Formation and Motion Blur Restoration via Multiple Capture, In ICASSP2001 conference, Salt Lake City, Utah, May 2001. [4] D. Yang, and A. El Gamal, Comparative Analysis of SNR for Image Sensors with Enhanced Dynamic Range, In Proceedings of the SPIE Electronic Imaging 99 conference, Vol. 3649, San Jose, CA, January 1999. [5] T. F. Knight, Design of an Integrated Optical Sensor with On-Chip Preprocessing. PhD thesis, MIT, 1983. [6] M. Sayag, Non-linear Photosite Response in CCD Imagers. U.S Patent No. 5,055,667, 1991. Filed 1990. [7] S. J. Decker, R. D. McGrath, K. Brehmer, and C. G. Sodini, A 256x256 CMOS imaging array with wide dynamic range pixels and column-parallel digital output. IEEE J. of Solid State Circuits, Vol. 33, pp. 2081-2091, Dec. 1998. [8] O. Yadid-Pecht and E. Fossum, Wide intrascene dynamic range CMOS APS using dual sampling. IEEE Trans. Electron Devices, vol. 44, pp. 1721-1723, Oct. 1997. [9] D. X. D. Yang, A. El Gamal, B. Fowler, and H. Tian, A 640x512 CMOS image sensor with ultrawide dynamic range oating-point pixel level ADC, IEEE Journal of Solid State Circuits, vol. 34, pp. 1821-1834, Dec. 1999.

ISSCC02

59

[10] S. Kleinfelder, S.H. Lim, X.Q. Liu and A. El Gamal, A 10,000 Frames/s 0.18 um CMOS Digital Pixel Sensor with Pixel-Level Memory, In Proceedings of the 2001 IEEE International Solid-State Circuits Conference, pp. 88-89, San Francisco, CA, February 2001. [11] S. K. Nayar and T. Mitsunaga, High dynamic range imaging: Spatially varying pixel exposures. http://www.cs.columbia.edu/CAVE/, March 2000. [12] V. Brajovic, T. Kanade, A sorting image sensor: an example of massively parallel intensity-to-time processing for low-latency computational sensors. Proceedings of the 1996 IEEE International Conference on Robotics and Automation, Minneapolis, MN, pp. 1638-1643. [13] E. Culurciello, R. Etienne-Cummings, and K. Boahen, High dynamic range, arbitrated address event representation digital imager. The 2001 IEEE International Symposium on Circuits and Systems, Vol. 2, pp. 505-508, 2001. [14] V. Brajovic, R. Miyagawa and T. Kanade, Temporal photoreception for adaptive dynamic range image sensing and encoding. Neural Networks, Vol. 11, pp. 1149-1158, 1998. [15] S. Kavadias, B. Dierickx, D. Scheer, A. Alaerts, D. Uwaerts, J. Bogaerts, A Logarithmic Response CMOS Image Sensor with On-Chip Calibration, IEEE Journal of Solid-State Circuits, Volume 35, Number 8, August 2000 [16] C. A. Mead and M. A. Mahowald, A Silicon Model of Early Visual Processing, Neural Networks, vol. 1, pp. 91-97, 1988. [17] C. A. Mead, Adaptive Retina, Analog VLSI Implementation of Neural Systems, C. Mead and M. Ismail, Eds., Boston: Kluwer Academic Pub., pp. 239-246, 1989.

Supplementary References:
ISSCC02 60

[18] M. Aggarwal, N. Ahuja, High dynamic range panoramic imaging. Proceedings of Eighth IEEE International Conference on Computer Vision, Vol. 1, pp. 2-9, 2001. [19] V. Brajovic, Sensory Computing. Conference 4109: Critical Technologies for the Future of Computing, SPIEs 45th Annual Meeting, SPIE, 2000. [20] S. Chen, R. Ginosar, Adaptive sensitivity CCD image sensor. Conference C: Signal Processing, Proceedings of the 12th IAPR International Conference on Pattern Recognition, Vol. 3, pp. 363-365, 1994. [21] T. Delbruck and C. A. Mead, Analog VLSI phototransduction. California Institute of Technology, CNS Memo No. 30, May 11, 1994. [22] T. Hamamoto, K. Aizawa, A computational image sensor with adaptive pixel-based integration time IEEE Journal of Solid-State Circuits, Vol. 36, Issue 4, pp. 580-585, April 2001. [23] M. Loose, K. Meier, J. Schemmel, A self-calibrating single-chip CMOS camera with logarithmic response. IEEE Journal of Solid-State Circuits, Vol. 36, Issue 4, pp. 586-596, April 2001. [24] C. A. Mead, A Sensitive Electronic Photoreceptor, In 1985 Chapel Hill Conference on VLSI, H. Fuchs, Ed., Rockville: Computer Science Press, pp. 463-471, 1985. [25] N. Ricquier and B. Dierickx, Pixel structure with logarithmic response for intelligent and exible imager architectures, Microelectronic Eng., vol. 19, pp. 631-634, 1992. [26] M. Schanz, C. Nitta, A. Bussmann, B.J. Hosticka, and R.K. Wertheimer, A high-dynamic-range CMOS image sensor for automotive applications. IEEE Journal of Solid-State Circuits, Vol. 35, Issue 7, pp. 932-938, July 2000. [27] O. Schrey, R. Hauschild, B.J. Hosticka, U. Iurgel, and M. Schwarz, A locally adaptive CMOS image sensor with 90 dB dynamic range. Solid-State Circuits Conference Digest of Technical Papers, pp. 310-311, 1999.
ISSCC02 61

[28] Y. Wang, S. Barna, S. Campbell, and E. Fossum, A High Dynamic Range CMOS Image Sensor. In Proceedings of 2001 IEEE Workshop on Charge-Coupled Devices and Advanced Image Sensors, pp. 137-140, June 2001. [29] O. Yadid-Pecht, Wide dynamic range sensors. Optical Engineering, Vol. 38, No. 10, pp. 1650-1660, October 1999. [30] Yadid-Pecht, O.; Belenky, A. Autoscaling CMOS APS with customized increase of dynamic range. Solid-State Circuits Conference: Digest of Technical Papers. pp. 100-101, 2001. [31] W. Yang, A wide-dynamic-range, low power photosensor array. IEEE ISSCC Vol. 37, 1994.

ISSCC02

62