A Method to Achieve High Dynamic Range in a CMOS Image Sensor Using Interleaved Row Readout Thomas Wocial1 Konstantin D. Stefanov2 William E. Martin1 John R. Barnes2

2025-04-27 0 0 931.3KB 14 页 10玖币

侵权投诉

A Method to Achieve High Dynamic Range in a CMOS Image

Sensor Using Interleaved Row Readout

Thomas Wocial1, Konstantin D. Stefanov2, William E. Martin1, John R. Barnes2,

Hugh R.A. Jones1

Abstract

We present a readout scheme for CMOS image sen-

sors that can be used to achieve arbitrarily high dy-

namic range (HDR) in principle. The linear full

well capacity (LFWC) in high signal regions was ex-

tended 50 times from 20 ke−to 984 ke−via an inter-

laced row-wise readout order, whilst the noise ﬂoor

remained unchanged in low signal regions, resulting

in a 34 dB increase in DR. The peak signal-to-noise

ratio (PSNR) is increased in a continuous fashion

from 43 dB to 60 dB. This was achieved by sum-

ming user-selected rows which were read out multiple

times. Centroiding uncertainties were lowered when

template-ﬁtting a projected pattern, compared to the

standard readout scheme. Example applications are

aimed at scientiﬁc imaging due to the linearity and

PSNR increase.

1 Introduction

In recent years, the CMOS image sensor (CIS) has

seen increased adoption by the astronomical com-

munity, particularly in time-domain applications [1].

First invented in 1993 [2], some advantages of the

active pixel architecture over charge-coupled devices

Manuscript received 13/08/2022; revised 12/09/2022; ac-

cepted 14/09/2022. Date of publication 06/10/2022; date of

current version 10/10/2022. TW, WM and HJ acknowledge

support from STFC grants ST/W507490/1, ST/W508020/1,

ST/P005667/1 and ST/R006598/1. This work was made pos-

sible through a Research, Enterprise and Scholarship Innova-

tion grant from the Open University, which funded both KDS

and JRB.

1TW, WM and HJ are with the Centre for Astrophysics

Research, University of Hertfordshire, College Lane, Hatﬁeld

AL10 9AB, UK (e-mail: t.wocial@herts.ac.uk).

2KDS and JRB are with the School of Physical Sci-

ences, Open University, MK7 6AA, UK (e-mail: kon-

stantin.stefanov@open.ac.uk).

Digital Object Identiﬁer: 10.1109/JSEN.2022.3211152

(CCDs) include 1) ability for faster readout, 2) read-

out from region of interest (ROI), 3) low power, 4)

low readout noise and 5) built-in anti-blooming.

A variety of techniques exist to extend the dy-

namic range (DR) of CMOS image sensors. These

have previously been classiﬁed into seven categories:

1) logarithmic pixel response, 2) combined linear and

logarithmic response, 3) well capacity adjustment, 4)

frequency based sensors, 5) time-to-saturation based

sensors, 6) global control of integration and 7) lo-

cal control over integration [3][4]. Other advances

that result in an increased DR include improved dark

current suppression, lower read noise and multi gain

readout [5].

The 4/5T pixel architecture is widely employed in

scientiﬁc imaging, but the selection of DR extension

techniques is limited. For instance, dynamic ranges

of 160 dB have been achieved in CIS using a 3T ar-

chitecture with integrated charge compensation pho-

todiode [6] and 141 dB (proposed) on 4T architecture

with a lateral overﬂow integration capacitor but both

suﬀer from non-linear pixel responses and signal-to-

noise ratio (SNR) curves. DR extension techniques

used for comparison to this work will apply to 4T

pixels and therefore fall under: local control over in-

tegration and global control over integration.

For the scientiﬁc requirement and operation of our

CIS, key assumptions are made about the scene to

be imaged: 1) there exists a desired minimum SNR

of the observation, hence minimum integration time,

2) the scene is spatially and temporally static, 3) it

is always desirable to achieve the maximum possible

SNR for any local region.

Self-reset pixels achieve HDR by locally controlling

integration time. The pixel is based on the 4T ar-

chitecture with additional circuitry to trigger a reset

signal when the voltage at the sense node matches

a reference corresponding to the full well capacity

(FWC). By counting the triggers and sampling the

arXiv:2210.04824v1 [physics.ins-det] 10 Oct 2022

residual voltage it is possible to reconstruct photo-

signals that would otherwise exceed the pixel FWC

for a given integration whilst preserving linearity. Re-

cent implementations include a 16×16 pixel on 20 µm

pitch at 121 dB DR [7] and 96×128 pixel on 25 µm

pitch at 125 dB DR [8]. Fill factors are 13.1% and

10% respectively. A key advantage is the continu-

ous SNR increase with photosignal, however the low

ﬁll factor, large pitch and small array sizes can be a

challenge for use in scientiﬁc imaging.

Multiple exposure is a commonly employed tech-

nique to increase DR without the need for additional

circuitry. Dual exposure involves acquiring two im-

ages at diﬀerent integration times Tlong and Tshort

with the DR extension equal to Tlong/Tshort [9][10].

There exists an SNR dip in regions corresponding to

saturation in the short exposure, as the photosignal

here is only sampled for a fraction of the total inte-

gration time. The use of multiple shorter integrations

can lower the resulting SNR dip [11].

Non-destructive readout (NDR) operates by sam-

pling the signal many times during the integration

time without aﬀecting the built up photocharge. The

CIS is read with up-the-ramp sampling, as widely

used on IR photodiode arrays for DR increase [12][13]

and for cosmic ray rejection [14].

Coded rolling shutter [15] works by spatially vary-

ing exposure per row, achieving a DR increase up

to the ratio of longest and shortest exposure times.

Image reconstruction is needed as vertical spatial in-

formation may be lost, however, due to the encoded

temporal information, high speed video and optical

ﬂow can be extracted [16][17]. Interleaved multiple

gain readout has been used to achieve a DR of 120 dB

[18]. Work in this area is limited by commercial CIS

devices often being addressable row-wise only, pixel-

wise coded exposure could be achieved with full X-Y

addressability [16].

Pixel-wise control of integration times builds on

row-wise coded exposure by the addition of control

over an additional spatial dimension. Recent ad-

vances in 3D stacked CIS devices haven enabled a

pixel-parallel architecture, typically based on Cu-Cu

interconnects between the sensing and logic layers

[19]. A similar method is to control exposure for a

block of pixels, as shown in [20][21]. In [22] a 512×512

array with 4.6 µm pitch on a 3D-IC achieves 127 dB

DR by combining dual conversion gain with time-

to-saturation detection. Key advantages are the high

spatial ﬁdelity achieved with a pixel-parallel readout,

however there is non-continuity in the SNR response

due to the three photosensing regimes.

In this paper a row-wise coded exposure scheme

is proposed which varies exposure locally depending

on the illumination, with multiple sub-exposures oc-

curring in said regions to exceed the SNR limit im-

posed by the FWC of the pixel. A hardware demon-

stration on a Teledyne e2v (Te2v) SIRIUS CIS115

[23] results in extended DR and peak signal-to-noise

ratio (PSNR). This is achieved thanks to the abil-

ity to randomly address and read out rows. A fu-

ture implementation of this technique is for actively

controlled spectroscopy [24], as both high signal cal-

ibration spectral lines and science spectral lines fall

on the focal plane [25][26]. Other applications may

include multi-ROI imaging at varying sub-exposures

and other areas, where high DR and SNR images are

of interest. The ability for CMOS image sensors to

operate in such way has been recognised by others

[27][28], but to our knowledge has not been imple-

mented in scientiﬁc imaging applications.

2 Dynamic range

A scene’s dynamic range is given by the upper and

lower limits of its luminance range. It is desirable for

a CIS to have a high dynamic range as this means it

can faithfully quantify both the high and low signal

regions in the scene. For a sensor with a linear full

well capacity of Qwell (e−), exposure time t(s), aver-

age dark current Idc (e−/s) and read noise variance

σ2

read (e2), the DR is the ratio of maximum (Imax)

and minimum (Imin) detectable photocurrents as in

equation 1. The maximum detectable signal is lim-

ited by Qwell minus the dark signal (which is sub-

tracted from a reference image taken in darkness),

whilst the noise ﬂoor (assuming correlated double

sampling and ﬁxed pattern noise correction) is de-

termined by the read noise and dark current signal

noise.

DR = 20 log10

Imax

Imin

= 20 log10

Qwell −Idct

pσ2

read +Idct(1)

The readout noise determines the noise ﬂoor of a

CIS. This value is almost independent of tempera-

ture and integration time. It is comprised of tran-

sistor noise from the source follower, ampliﬁer and

analogue-to-digital converter (ADC) noise. Quanti-

zation noise can contribute to read noise if the sig-

nal is sampled with low resolution. To determine

the readout noise we calculate the standard devia-

tion of each pixel using a set of bias frames [29]. The

input-referred noise is often given in equivalent noise

charge, expressed in e−RMS.

Dark current occurs as a result of thermal excita-

tion of electron-hole pairs in silicon. The sensor used

in this paper, the CIS115, has a dark current of 20

e−/pix/s at 293 K, which halves for every temper-

ature reduction of 5.5 K [23]. If the noise ﬂoor is

dominated by dark current, using shorter exposures

or cooling the CIS can increase the DR.

3 Method

3.1 CIS115

The sensor used is a CIS115 from Te2v, featuring an

array of 2000 ×1504 pixels on 7 µm pitch [23] (see

Table 1 for measured BSI variant speciﬁcations). It

employs the 4T pixel architecture and is fabricated

using a 0.18 µm CIS process. The model used in this

paper is a front side illuminated variant. A back side

illuminated variant has been adopted for the JANUS

instrument on JUICE [30][31]. The sensor is divided

into four blocks of 376 columns which are read out

simultaneously and in parallel. Each pixel transfers

the reset and signal levels to a storage buﬀer that

allows for correlated double sampling (CDS) to be

performed. The CIS115 operates in rolling shutter

mode meaning integration time is simultaneous for

all pixels in a row. All control signals are generated

externally.

Unit CIS115

Active rows 2000

Active columns 1504

Pixel size µm 7.0

Non-linearity ±% 3

Mean read noise e−5

Peak linear charge e−/pix 27,000

Saturation charge e−/pix 33,000

Dynamic range dB 74.6

Dark current e−/pix/s 20 (at 293 K)

DSNU e−/pix/s 12 (at 293 K)

Table 1: Speciﬁcations for the CIS115 [23]

3.2 Row-wise HDR readout scheme

The scheme demonstrated in this paper seeks to ad-

dress key traits for scientiﬁc imaging: pixel linearity

and continuous SNR increase with photosignal. For

a maximum per pixel photocurrent per row I(M), a

total integration time t, with row-wise control over

exposure, the scheme is as follows: If the FWC is

reached or exceeded in t, perform Nsub-exposures

with a read and reset sample such that equation 2 is

satisﬁed. N rows are summed (stacked) in software.

N=dI(M)×t/F W Ce(2)

By doing so, the noise ﬂoor is kept at a minimum

on a per-row basis and the PSNR is increased by

sampling the maximum detectable photosignal mul-

tiple times. We implement a simpliﬁed version of this

scheme with one region that can reconstruct photo-

signals 50 times greater than the FWC per integra-

tion.

The CIS115 is mounted on a control PCB, with

sensor interfacing achieved via a National Instru-

ments PXIe-7856R FPGA card. A custom LabVIEW

GUI was developed at the Open University. In stan-

dard conﬁguration the minimum row readout time,

Trow , is 412.5 µs so the readout time for the whole

array once is 825 ms. During readout the analogue

sensor outputs are digitised by four 16-bit ADCs in

the PXIe-7856R card with the digital values stored

as signed 32 bit integers. In this demonstration HDR

is achieved by sorting the rows (MT otal) into two

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

AMethodtoAchieveHighDynamicRangeinaCMOSImageSensorUsingInterleavedRowReadoutThomasWocial1,KonstantinD.Stefanov2,WilliamE.Martin1,JohnR.Barnes2,HughR.A.Jones1AbstractWepresentareadoutschemeforCMOSimagesen-sorsthatcanbeusedtoachievearbitrarilyhighdy-namicrange(HDR)inprinciple.Thelinearfullwellcapacity...

展开>> 收起<<

A Method to Achieve High Dynamic Range in a CMOS Image Sensor Using Interleaved Row Readout Thomas Wocial1 Konstantin D. Stefanov2 William E. Martin1 John R. Barnes2.pdf

共14页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

A Method to Achieve High Dynamic Range in a CMOS Image Sensor Using Interleaved Row Readout Thomas Wocial1 Konstantin D. Stefanov2 William E. Martin1 John R. Barnes2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: