Documente Academic
Documente Profesional
Documente Cultură
com
Abstract
In the 3eld of interferometric metrology the use of high resolution CCD sensors with 1024 × 1024 to 2048 × 2048 pixels is predominant.
Due to special features (e.g. random pixel access, characteristic curve) CMOS sensors with similar resolution can be an interesting
alternative. We compare some characteristics of both sensor types that are important for interferometry and demonstrate two exemplary
applications that are only possible by using CMOS cameras.
? 2003 Elsevier Ltd. All rights reserved.
0030-3992/03/$ - see front matter ? 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0030-3992(03)00078-1
588 H. Helmers, M. Schellenberg / Optics & Laser Technology 35 (2003) 587 – 595
Table 1
Characteristic data of the investigated high resolution CCD and CMOS cameras
1024 × 1024 or more, a comparable pixel size, and a digital poral noise of the camera signal. First we will de3ne both
LVDS or RS 422 interface. The protective glass in front of quantities.
the sensors was removed in order to avoid parasitic inter- Let g be the grey value (given in digital numbers, DN)
ferences. Characteristic data of the cameras (manufacturer’s of a pixel after exposure, for which we can write
indications) are given in Table 1. In the following chapters
the practical importance of some of these parameters will g = g E + gD ; (1)
be discussed.
gE is the grey value caused by the exposure E = It with I
being the light intensity. gE is proportional to NE , the number
3. Set-up for measurement of camera characteristics of electrons generated by the exposure. The proportionality
factor is given by the node sensitivity S and the resolution
For measurement of characteristic data of the cameras the and adjustment of the analogue-to-digital converter (ADC).
sensors were homogeneously illuminated. We used a volt- gD is the grey value for the darkened sensor at same t. gD
age regulated halogen lamp or a square array of 16 current is caused by Ni , by the oDset of the ampli3ers and the ADC
regulated high power LEDs (Sharp GL0ZJ042B0S) as light and by the electronic noise of the readout circuit. gE and
source. By means of a set of lenses and ground glasses a gD are statistically independent. Therefore the total temporal
homogeneous light 3eld was achieved at a distance of sev- noise (standard deviation) of g is given by
eral 10 cm behind the light source (uniformity better than
1% over the investigated sensor region). This simple linear = E2 + D2 ; (2)
set-up has an advantage compared to a set-up with an inte-
grating sphere, because almost vertical light incidence onto E is the noise (standard deviation) of gE caused by photon
the sensor is achieved. This ensures an optimal function of shot noise. For abbreviation, we will call E photon shot
the micro-lenses which are used in several sensors in order noise in the following text. D is the dark noise, i.e. the
to increase the 3ll factor. An interference 3lter (T = 628 nm, standard deviation of gD . It is caused by the reset noise r
FWHM = 7 nm) was used to limit the spectral bandwidth of the sense node, by the electronic noise (thermal noise) e
of the light. The light intensity I was measured with a cal- of the readout circuit, and by the shot noise i of the dark
ibrated photodiode with linear response over the required current i. r and e are independent of the exposure time t.
√
intensity range. E und i increase with t, because NE ∼ t and Ni ∼ t and
both quantities obey Poisson statistics. Therefore:
4. Dark current √ √ √
E ∼ NE ∼ t and i ∼ Ni ∼ t: (3)
The number of electrons per pixel generated by the dark For an ADC with a resolution of 8 bit or more the quantiza-
current i, Ni , doubles with an increase of the sensor temper- tion noise can be neglected compared to E and D . There-
ature T by about (5 –8)◦ C [9]. The data given in Table 1 fore it is not further discussed in the following.
were assessed on the basis of manufacturer’s indications and The total temporal noise as a function of the expo-
are valid for T ≈ 35◦ C, a typical value for the tempera- sure E is measured under homogeneous illumination of the
ture of many sensors (not cameras) during operation. As ex- sensor and constant exposure time t, while D is measured
pected [7,8,12], for the CMOS sensors i is about one order with darkened sensor and same t. Both quantities (as well as
of magnitude higher than for CCD sensors. In the 3eld of E ) are determined from the subtraction image of two im-
interferometry, where short exposure times or pulse expo- ages in order to eliminate the inCuence of remaining spatial
sure are typical, this is not a severe problem. Let us assume variations in the illumination and of the pattern noise (see
an exposure time of t = 20 ms. Then, for the cameras listed appendix).
in Table 1, the grey value caused by Ni does not exceed ap- Figs. 2 and 3 show the total temporal noise as a func-
proximately 0.5% of the maximum grey value for CCD and tion of GE = G − GD , where G, GE , and GD are the mean
CMOS sensors. This grey value can be measured once by values taken over m pixels (m ¿ 104 ) of the corresponding
recording a dark image which can then be subtracted from quantities given in Eq. (1), for diDerent CCD and CMOS
the images taken later. The increasing 3lling of the pixels cameras (photon transfer curves). For all exposure times in-
by the dark current would only be a problem for long ex- vestigated (t 6 1000 ms) is independent of t for all sen-
posure times in the range of some seconds. More important sors. From that we can conclude that the inCuence of the
than the dark current is its noise, which will be discussed in dark current on and therefore on D is negligible. There-
the next chapter. fore, D is primarily caused by the reset noise r and by the
electronic noise e . The shot noise of the dark current, i ,
5. Dark current shot noise and electronic noise plays only a subordinate role.
As shown in Figs. 2 and 3, for several cameras
In this chapter we will discuss the inCuence of the dark (MV-D1024K, TM1001/1010) the total temporal noise
current shot noise and the electronic noise on the total tem- decreases in the proximity of the maximum mean grey
590 H. Helmers, M. Schellenberg / Optics & Laser Technology 35 (2003) 587 – 595
4
C4742-95, 10 Bit TM1010, 10 Bit to the total noise . If NE electrons are √ generated by the
2 exposure, then E is proportional to NE . Therefore we
de3ne an eDective SNReD , which takes the photon shot noise
1 TM1001, 8 Bit into account:
0.5 NE
8 32 128 512 2048 SNReD : = √ : (4)
ND + N E
GE / DN
In case of pixel saturation we have NE + ND = fwc.
Fig. 2. Total temporal noise vs. GE for diDerent CCD cameras and Furthermore, for short exposure times (t1 s) ND NE and
diDerent exposure times t (DN: digital number). t = (64; 32; 16) ms for
ES 4.0, t = (1000; 500; 56; 36) ms for C4742-95, t = (16; 8; 4; 2) ms for
thus fwc ≈ NE . The best eDective SNR that can be achieved
TM1010, and t = (8; 4; 2; 1) ms for TM1001. The arrow below the sym- under these assumptions is therefore SNReD ≈ fwc1=2 (Eq.
bol line indicates increasing t. The lines are used to indicate related (4)). In other words, SNReD = 1024 (10 Bit) would require
measurements. fwc ≈ 106 electrons and SNReD = 4096 (12 Bit) would
require fwc ≈ 1:7 × 107 electrons. For the sensors listed in
4 Table 1, the actual full well capacities and therefore SNReD
are considerably smaller. Therefore, for these cameras an
t
ADC resolution of 10 bit or 12 bit only provides a better
2
resolution of the noise.
MV-D1024K, 10 Bit With increasing charge within a pixel the probability in-
creases that some electrons hop out oD the potential well of a
/ DN
1
pixel. This leads to a small nonlinearity of the characteristic
curve in the region of large grey values. In order to prevent
0.5
MV-D1024K, 8 Bit
this nonlinearity, some camera manufacturers adjust the am-
pli3cation and the ADC in a way that the maximum grey
value is already achieved at (ND + NE ) ¡ fwc. Then the
0.25
1 4 16 64 256 1024 4096 eDective full well capacity of a camera, fwceD , is smaller
GE / DN than the fwc of the sensor and therefore the eDective SNR
is reduced, too.
Fig. 3. Total temporal noise vs. GE for two CMOS cameras and diDerent A further reduction of fwceD is caused by an additional
exposure times t. t = (30; 20; 10) ms for MV-D1024K, 8 Bit and 10 Bit. electronic ampli3cation of the sensor signal by a factor A. In
The arrow below the symbol line indicates increasing t.
this case, fwc√eD is reduced by the same factor A and SNReD
by the factor A.
value. Furthermore, it can be recognized that (in DN) fwceD can be determined from a diagram of E vs. GE
is predominantly smaller for the CMOS cameras than for (for the calculation of E see appendix). Let us 3rst discuss
the CCD cameras. Both eDects are caused by the inCuence CCD cameras with linear characteristic curves. Let K be
of the photon shot noise E , which makes a considerable the ampli3cation constant of the camera which de3nes the
contribution to the total noise, and which will be discussed relation between the number N of electrons within a pixel
in the next chapter. and the corresponding grey value G, i.e.
G = GE + GD = KN = K(NE + ND ) (5)
6. Photon shot noise and signal-to-noise ratio (SNR)
with DN=e being the unit of K. Then it follows for GE :
Some camera manufacturers indicate the resolu-
tion of the ADC in bit as “dynamic” of a digital GE = G − GD = KNE (6)
camera. This quantity has only little to do with the and for the photon shot noise E :
achievable signal-to-noise ratio SNR = GE =. Fig. 2
√ √
clearly shows that, e.g., the SNR of the 12 bit camera ES E = K NE = K GE : (7)
4.0 (Kodak) and the 10 bit camera C4742-95 (Hamamatsu)
is only slightly higher than the SNR of the 8 bit camera If E is plotted vs. GE on a double logarithmic scale, we
TM1001 (Pulnix). get outside the region where saturation eDects are already
H. Helmers, M. Schellenberg / Optics & Laser Technology 35 (2003) 587 – 595 591
32 2
fwceff / ke
16 ES 4.0 41.6 ± 0.4
TM 1010 51.1 ± 0.8 1 MV-D1024K, 10 Bit
8 TM 1001 26.8 ± 0.3 fwceff = (168 ± 3) ke
C 4742-95 32.9 ± 0.3
0.5
E / DN
E / DN
2 0.25
MV-D1024K, 8 Bit
1 fwceff = (163 ± 3) ke
0.125
0.5
0.25 0.0625
8 32 128 512 2048 2 4 8 16 32 64 128 256 512 1024
GE / DN GE / DN
Fig. 4. Photon shot noise E vs. GE with lines of regression for the Fig. 5. Photon shot noise E vs. GE for two CMOS cameras. The straight
determination of fwceD for diDerent CCD cameras. lines represent the theoretically expected course of E vs. GE on the basis
of Eq. (11).
√
noticeable, a straight line with slope 12 and oDset log( K)
in Eq. (9):
from which K can be determined. The maximum mean grey
√
value Gmax is reached when N = fwceD electrons are accu- E = bkNEb−1 NE : (10)
mulated, so that we 3nally get (Eq. (6)):
After insertion of Eq. (9) it follows:
Gmax
fwceD = : (8)
K E = bk 1=2b (GE )1−(1=2b) : (11)
Fig. 4 shows the results for the investigated CCD cameras. If E is plotted vs. GE on a double logarithmic scale
Two of the determined values for the fwceD of the cameras we get a straight line with slope (1 − 1=(2b)) and oDset
are in good agreement with the fwc of the sensors (ES4.0, log(bk 1=2b ). These data are shown in Fig. 5. The parameter
TM1010) as listed in Table 1. Two other values show con- b can be determined from a nonlinear curve 3t applied to
siderable deviations. For the camera TM 1001 the deviation the characteristic curve (Eq. (9)), and k from a nonlinear
has the expected behaviour (fwceD ¡ fwc) and is caused curve 3t applied to Eq. (11).
by the individual adjustment of the camera parameters by From k and b the quantity fwceD can be estimated (Eq.
the manufacturer, as mentioned above. The deviation for the (9)) under the approximations NE ND (i.e. fwceD ≈ NE )
camera C4742-95 cannot be explained in this way, because and that Eq. (9) is also valid in the proximity of Gmax so
here fwceD ¿ fwc, which is nonsense from a physical point that
of view. For a 12 bit version of the same camera we found 1=b
fwceD = (23 ± 1) ke, which in turn is larger than fwc. The Gmax
fwceD ≈ : (12)
reason for this behaviour could possibly be a low-pass 3l- k
tering of the sensor signal which smoothes the Cuctuations
caused by the photon shot noise. We were in close contact Under these assumptions we get values for fwceD of the
with the manufacturer in order to clarify this question. How- cameras that are ∼15% smaller than the fwc values for the
ever, the problem could not be solved till now. sensors (see Fig. 5 and Table 1). The diDerences are pre-
For CMOS cameras with nonlinear characteristic curves sumably caused by the adjustment of the camera parameters
(see Chapter 8) we get other relations than for CCD cam- by the manufacturer and by the mentioned approximations.
eras with linear characteristic curves. For the CMOS cam- A comparison between the CCD cameras and the CMOS
eras investigated in this work the characteristic curve can cameras investigated in this work shows that the CMOS
approximately be described by a power function, as long cameras have a higher eDective fwceD . Thereby the inCu-
as we concentrate on the region NE ND und NE ¡ fwceD ence of the photon shot noise on the SNR is reduced. This
(i.e. below the region, where saturation eDects occur). In explains, why the total temporal noise (in DN) of the
this case the relation between GE and NE or between GE CMOS cameras is smaller that that of the CCD cameras.
and the exposure E can be approximated by For some cameras (MV-D1024K, TM1001/1010) the to-
tal temporal noise and the photon shot noise E decrease
GE = kNEb = aE b (9) in the proximity of the maximum mean grey value (see
Figs. 2–5). This is caused by saturation of the ampli3ers
with k; b and a being constants. The photon shot noise and by the increasing 3lling of the pixels so that the proba-
E can be found by applying the error propagation bility increases that electrons hop out oD the potential well.
592 H. Helmers, M. Schellenberg / Optics & Laser Technology 35 (2003) 587 – 595
Due to both eDects those Cuctuations of NE are cut oD, that 1024
lie above fwceD , whereby a lower noise is suggested. CCD C4742-95 / 10 Bit
CCD TM1001 / 8 Bit
768
7. Pattern noise
GE / DN
512
CMOS MV-D1024K / 10 Bit
In contrast to CCD sensors CMOS sensors are active pixel CMOS MV-D1024K / 8 Bit
sensors (APS) where every pixel has its own sense node and 256
1024 9. Sensitivity
"Linear"
Using the data in Fig. 7 the sensitivity of the cameras
768
(here at 628 nm) can be compared. In order to avoid the
inCuence of saturation eDects in the proximity of Gmax ,
GE / DN
®
LINLOG
512 we de3ne as comparative quantity the exposure E1=2 for
which G = Gmax =2. For the two CCD cameras we found
E1=2 ≈ 0:025 J=cm2 (C4742-95) and E1=2 ≈ 0:023 J=cm2
256
(TM1001). For the CMOS cameras MV-D1024K we found
E1=2 ≈ 0:12 J=cm2 (10 bit) and E1=2 ≈ 0:088 J=cm2
0 (8 bit). The main reason why E1=2 for the CCD cameras is
0.0 0.1 0.2 0.3 0.4 0.5 0.6
-2
about 4 –15 times smaller than for the CMOS cameras is the
E / µJ cm
diDerent fwc of the sensors. For the CCD camera TM1001,
Fig. 8. Characteristic curve of the 10 Bit CMOS camera MV-D1024K in e.g., the fwceD is about 6.3 times smaller than the fwceD of
the “linear” and in one of many possible LINLOGJ -modes of operation. the 10 bit CMOS camera MV-D1024K. Consequently, con-
siderably less absorbed photons are necessary for pixel sat-
uration, suggesting a much better sensitivity. On the other
hand, starting from the required exposure ENE to generate
parameter set p. For comparison purposes the characteristic the same number NE of electrons within a pixel, we get
curve of the same camera in the so-called “linear” mode of (for NE = 25 ke) ENE ≈ 0:044 J=cm2 for the CCD cam-
operation (Fig. 7) is shown in the diagram as well. era TM1001 and ENE ≈ 0:049 J=cm2 for the 10 bit CMOS
The nonlinearity of the characteristic curve can be used camera MV-D1024K. Under this point of view the sensi-
for interferometric measurements at objects with large tivity of both cameras is similar. This can be understood
brightness variations. If CCD cameras are used in such by comparing the quantum eNciencies of both sensors.
situations, saturation eDects and blooming may occur. Both Indeed, including the 3ll factor ≈ 0:27 for the CMOS
eDects can be prevented by using a CMOS camera with a sensor and ≈ 0:26 for the CCD sensor (manufacturer’ in-
nonlinear characteristic curve. As an example, Fig. 9 shows dications). That is, the smaller 3ll factor of the CMOS sen-
the result of an ESPI deformation measurement. The object sor (see Table 1) is compensated by its better absorption
was a PC graphic board and the deformation was caused characteristics.
by temperature variations. Although a 12 bit CCD camera As a result we can state that an exact de3nition of the
(Kodak ES 4.0) was used, the high reCectivity of some comparative quantity is important, if the sensitivity of cam-
components on the board caused pixel saturation so that no eras shall be compared.
deformation fringes were visible on these components. This
saturation is prevented by the nonlinear characteristic curve
of the 8 bit CMOS camera (Photonfocus MV D-1024k, 10. Regions of interest (ROI)
LINLOGJ -mode) so that now deformation fringes are
clearly visible in the formerly saturated regions. Due to the random pixel access the readout process of
For a quantitative analysis of deformation fringes using CMOS sensors can be limited to ROIs. Thereby a high frame
phase shifting techniques the nonlinearity of the character- rate for these ROIs is achievable. This enables, e.g., fast
istic curve can be corrected with a suitable look-up-table. object deformations to be tracked with ESPI. Fig. 10 shows
Fig. 9. Results of ESPI deformation measurements with a 12 bit CCD camera (left, linear characteristic curve) and a 8 bit CMOS camera (right, nonlinear
characteristic curve). The white arrows indicate locations of saturation.
594 H. Helmers, M. Schellenberg / Optics & Laser Technology 35 (2003) 587 – 595
Fig. 10. Results of ESPI deformation measurements using a 8 bit CMOS camera with high frame rate. ROI = 128 × 128 pixels, t = 0:5 ms. Time
diDerence between the images 1:11 ms (frame rate 900 fps).
11. Summary The standard deviation WD of the subtraction image D1 −D2
is then given by
Some physical characteristics of interline-transfer CCD √
WD = D1 2 + 2 = 2D : (A.2)
cameras and integrating CMOS cameras with global shut- D2
ter are compared on the basis of experimental data. With
regard to spatial resolution, temporal noise, and sensitivity Analogously we can write for two images P1 und P2 taken
both types of cameras are comparable. DiDerences between with homogeneously illuminated sensor and same exposure
the cameras occur in the pattern noise and the characteris- time t:
tic curve. The pattern noise of the CMOS cameras is con-
siderably higher than that of the CCD cameras. It can be GP1 ≈ GP2 : =G; P1 ≈ P2 : = (A.3)
minimized by hardware (correlated double sampling) or with GPj = GEj + GDj (j = 1; 2). Furthermore we have
by software (using correction matrices). The characteristic
curves of the CCD cameras are linear, while they are non- GE1 ≈ GE2 : =GE ; E1 ≈ E2 : =E : (A.4)
linear for the CMOS cameras, even in the “linear” mode
of operation. The characteristic curve of the CMOS camera In analogy to Eq. (A.2) and using Eq. (2) the standard
can be changed to a more logarithmic course by adjusting deviation WP of the subtraction image P1 −P2 can be written
electronic parameters (LINLOGJ principle). This allows as
the recording of light 3elds with large brightness diDerences √ √
without pixel saturation and blooming. WP = 2 = 2 E2 + D2 : (A.5)
Two applications of CMOS cameras in ESPI are
demonstrated. In one case we showed that the nonlinear The total temporal noise , the dark noise D , and the photon
characteristic curve prevents saturation eDects in defor- shot noise E can therefore be calculated from the measured
mation measurements for an object with strongly varying quantities WP and WD by
reCectivity. In the other case we used the possibility of √ √
fast readout of small regions of interest in order to enable = WP = 2; D = WD = 2;
the tracking of a fast deformation process with high frame √
E = WP 2 − 2 = 2: (A.6)
rate. WD
H. Helmers, M. Schellenberg / Optics & Laser Technology 35 (2003) 587 – 595 595