pixels for multiple-tier processes

basketontarioΗλεκτρονική - Συσκευές

2 Νοε 2013 (πριν από 3 χρόνια και 9 μήνες)

71 εμφανίσεις

Bio
-
inspired design:
nonlinear digital
pixels
for multiple
-
tier
processes
(invited paper)

SPIE
Nano
/Bio/Info
-
Tech

Sensors and Systems

March 2013

O.
Skorka
, A.
Mahmoodi
,
J
.
Li,
and D.
Joseph

Electrical and Computer Engineering

University of Alberta, Edmonton, AB, Canada

Outline


Introduction


Bio
-
inspired design


Nonlinear response


Digital pixels


Multiple
-
tier processes


Conclusion

2

Introduction


Carver
Mead, who cofounded
Foveon

in 1997, was a
pioneer for bio
-
inspired electronics in the 1980s.


His work on the “silicon retina” continues to inspire
researchers in the image
sensor
community.


Yet, a
design inspired
by biological systems
too literally
does not guarantee comparable
functionality
.


The approach presented here is more concerned with
bio
-
inspired performance than structure or form.


We use the human eye’s capabilities as benchmarks to
motivate and direct research on image sensors, while
leveraging existing and emerging technologies.

3

Bio
-
inspired
design

Parameter

Type

1.

Power consumption

PwC

2.

Visual field

VF

Geometric

3.

Spatial resolution

SR

4.

Temporal resolution

TR

5.

Peak signal
-
to
-
noise
-
and
-
distortion

ratio

PSNDR

Signal

and

noise
power

6.

Dark limit

DL

7.

Dynamic

range

DR

4

Bio
-
inspired
design


Assuming an ideal lens, image sensors surpass the human eye
in the marked quadrants on both parameters indicated.

5

Bio
-
inspired
design


Of the 24 image sensors (2000

2010) surveyed, DR and DL
tended to be the most limiting factors; these limitations are
characteristic features of commercial image sensors.

6

Nonlinear
response

DR

SNDR

Logarithmic sensor

Wide DR

Low PSNDR

Linear sensor

Narrow DR

High PSNDR

Images © IMS
Chips
http
://www.ims
-
chips.de/

7

Nonlinear
response


With linear
sensors,
photodiode
capacitance is
first
charged at
the beginning of each
frame, and then
discharged
by
photocurrent during exposure.


With logarithmic sensors, response
is achieved
via
a
CMOS
transistor (
T
pd
) operating
in
sub
-
threshold.

Linear sensor

Logarithmic sensor

8

Nonlinear
response


T
wo paths to achieve wide DR and high PSNDR:


Improve DR of high
-
PSNDR linear sensors;


Improve SNDR of wide
-
DR nonlinear sensors.


SNDR is affected by two types of noise:


Temporal
noise, i.e., time
-
varying noise;


Spatial distortion, i.e., fixed pattern
noise (FPN
).


Our team developed new methods for FPN correction,
of nonlinear sensors, to the limit of temporal noise.


The methods, which are computationally efficient, are
suitable for both still
-
image and video applications.

9

Nonlinear
response


The image shown was taken with logarithmic CMOS
active pixel sensor (APS)

arrays designed in our lab.


Initial calibration of the sensor array is required.

Original image

Corrected

image

10

Digital
pixels

11


FPN is corrected to
the limit of temporal
noise with methods
we developed.


However, PSNDR
of
logarithmic CMOS APS
arrays
remains low
because it is limited
by PSNR.


No logarithmic CMOS APS array has achieved a PSNDR
higher than that of the human eye.

Digital pixels


Linear CMOS APS
arrays are integrating designs:


Finite
-
duration integration approximates a
first
-
order
low
-
pass filter (LPF) with a narrow bandwidth (BW);


The LPF reduces
temporal noise
, resulting in
high PSNR,
which enables high PSNDR and high image quality.


Logarithmic CMOS APS
arrays are
non
-
integrating:


Consequently,
each pixel sensor has
a wide BW;


The wide noise spectrum results
in low PSNDR
.


With
CMOS
APS
arrays, data conversion is
done
at chip
or column
level;
pixel
-
level digitization enables noise
filtering and protection from further noise.

12

Digital
pixels


We
designed and tested
logarithmic
CMOS
digital pixel
sensor (DPS)
arrays
with
fully
-
integrated
ΔΣ

ADCs.


They demonstrated a 43 dB PSNDR at video rates,
surpassing the human eye’s PSNDR of 36 dB.

13

Pixel layout

Sample image

Multiple
-
tier processes


Also, the DPS array’s SR is low because the
ΔΣ

ADCs
each require many transistors; in other words, pixel
size is too large for standard optical imaging.

14


DL of the DPS array is
comparable to that of
standard CMOS APS
(and CCD) arrays.


But it is two orders of
magnitude worse than
the human eye’s DL
(
colour

vision).

Multiple
-
tier
processes


Vertical
integration of heterogeneous tiers allows an
optimized process to be used with each one
.


F
abrication of ADC circuits in a nanoscale process
facilitates DPS arrays with higher SR.

15


SR depends also on
integration technology,
which is improving.


Photodetector

optimization can
improve DL.

Multiple
-
tier processes


We
designed and tested
logarithmic vertically
-
integrated (VI) CMOS APS
arrays that were assembled
by flip
-
chip bonding
.


Designs are composed of
CMOS and
photodetector

dies, which employ
hydrogenated
amorphous
silicon

detectors.

16


The VI
-
CMOS APS array demonstrated a DL that is an
order of magnitude better than that of typical CMOS
APS (and CCD) sensor arrays.

Multiple
-
tier processes


Sensor 25, the CMOS APS array, has wide
DR
but low PSNDR.


Sensor 27, the VI
-
CMOS APS array, has low DL; it is compatible
with Sensor 26, the CMOS DPS array, which has high PSNDR.


We expect superior performance with nonlinear digital pixels
in multiple
-
tier processes, i.e., with VI
-
CMOS DPS arrays.

17

Conclusion


The approach presented here for image sensor design
is inspired by the performance of the corresponding
biological system rather than by its structure
.


Compared to the human eye, DR
and DL are
the
most
limiting factors of
conventional
image
sensors.


FPN of logarithmic CMOS APS arrays, which easily
achieve wide DR, can be corrected effectively.


With CMOS DPS arrays,
in
-
pixel
ΔΣ

ADCs are used to
filter temporal noise and achieve high PSNDR
.


To achieve low DL, and to pursue high SR, multiple
-
tier
processes, or VI
-
CMOS technology, is investigated.

18

Acknowledgments


The authors would
like to thank:


Dr. Kamal
Ranaweera
;


Dr.
Jianzeng

Xu
;


Dr. Glen
Fitzpatrick;


NSERC;


Alberta Innovates

Technology Futures;


CMC Microsystems;


Micralyne
.

19