Quantitative Study of High Dynamic Quantitative Study of High Dynamic Range Image Sensor ArchitecturesRange Image Sensor Architectures

Sam Kavusi and Abbas El GamalDepartment of Electrical Engineering

Stanford University

Scene DR > Image Sensor DRScene DR > Image Sensor DR

ImageScene

Courtesy Pixim Corporation

DR Extension MethodsDR Extension MethodsSeveral methods have been developed to extend image sensor DRDeep submicron CMOS image sensor processes [Wuu ‘01] and vertical integration [Kozlowski ‘02] enable implementation of high DR, high fidelity methods

This work studies four such schemes:Time-to-saturation [Lule ‘99, Stoppa ‘02]Multiple-capture [Yang ‘99, Kleinfelder ‘01, Bidermann ‘03]Asynchronous self-reset with multiple capture [Liu ‘03]Synchronous self-reset with residue readout [Bermak ‘02, Rhee ‘03]

SNR as Fidelity MeasureSNR as Fidelity MeasureHow should these methods be compared?

Methods can be partially compared based on their SNR [Yang ’99, El Gamal ’02]:

Some schemes extend DR but at expense of SNR

This work: Quantify SNR for the four schemesConsider non-idealities due to implementation, especially due to pixel area constraint

Conventional (Reference) SensorConventional (Reference) SensorCCD, APS, CTIA

Timetint

v(t)Vmax

Vmin

iph(A)

TimeTime--toto--Saturation Saturation [Lule ‘99, Stoppa ‘02]

Pixel

Time

v(t)

tint

V1

tsat

Vmax

TimeTime--toto--Saturation Saturation [Lule ‘99, Stoppa ‘02]

Pixel

Similar to reference sensorand SNR limited by:

tsat inaccuracy due to comparator noise/delay, time-ramp inaccuracy, kTC of CT-Ref (inaccuracy in tsat measured by σsat)

SNRSNR

iph(A)

MultipleMultiple--CaptureCapture[Yang ‘99, Kleinfelder ’01, Bidermann ’03]

Pixel

v(t)

tint

Vmax

tcapt Timejtcapt

Vlast

MultipleMultiple--CaptureCapture[Yang ‘99, Kleinfelder ’01, Bidermann ’03]

Pixel

v(t)

tint

Vmax

tcapt

V1

Time

MultipleMultiple--CaptureCapture[Yang ‘99, Kleinfelder ‘01, Bidermann ‘03]

Pixel

Pixel-level ADC resolution limited (ADC-ramp, speed/resolution trade-off)Digital weighted averaging is possible [Liu ’03] Increasing tint is possible by motion blur prevention [Liu ’03]

Capture time is very accurate

SNRSNR

iph(A)

Asynchronous SelfAsynchronous Self--Reset Reset [Liu ’02]

Pixel

Time

v(t)

tint

Vmax

tcapt

Asynchronous SelfAsynchronous Self--Reset Reset [Liu ’02]

SNR better than MC but with higher DSP cost:ADC resolution limited (ADC-ramp, speed/resolution trade-off)Digital weighted averaging can compensate for itIncreasing tint is possible by motion blur prevention [Liu ’03]Self-reset accuracy is relaxed

SNR increasing but limited by gain FPN

Pixel

SNRSNR

iph(A)

Synchronous SelfSynchronous Self--Reset with Reset with Residue ReadoutResidue Readout

[Bermak ‘02, Rhee ’03]

Pixel

Time

v(t)

tint

Vmax

tclk Time

v(t)

tint

Vmax

tclk

Synchronous SelfSynchronous Self--Reset with Reset with Residue ReadoutResidue Readout

[Bermak ‘02, Rhee ‘03]

Pixel

Reset noise and ADC resolutionimax and SNR at high end limited by:

Comparator and self-reset offset accumulation resulting in gain FPN (inaccuracy measured by σoffset)Underestimation of iph due to signal saturation

SNRSNR

iph(A)

SummarySummary

Method imin imax SNRPixel

MismatchesEffect

DSP Power

TTS = ++ – High Low

Moderate

High

Moderate

MC – /= + + Moderate

Comparator

Readout/DSP

Readout/DSPAsync – /= + ++ Moderate

Sync – – +++ – – High Comparator/digital circuits