is clearly no,
ss. For frame-
priate. As an
ize of 7.4 um
ited with the
sor, however,
d for imaging
1-limited lens
[6]
(0)
Xf d = 3.8 um
t, show many
s
1995 2000
ixels. Record
1 of the date
1983, that the
he laboratory.
1995 2000
n an image
plotted as a
2-24 1995
producing such large CCDs is the pricing policy of the manufacturers. Consider, for example, the sensor price and grade structure
announced by Loral Fairchild in mid-1994 for their 2048x2048 CCD 442A: Their best imagers, termed grade 1, suffer from about
200 bad pixels and typically three bad columns per CCD. Nevertheless, such an image sensor costs $ 15,000. Lower quality CCDs
are available at lower prices. A grade 4 image sensor costs just $ 2000. However, about 20,000 bad pixels must be tolerated, and
typically about 100 columns will be bad [13]. There are several reasons for a "bad pixel" or a "bad column", among them oxide
pinholes (small holes in the otherwise insulating oxide, through which current could pass from one layer to the next), local silicon
contaminations or defects near the surface (causing regions of increased dark current, so-called "dark spikes" or "hot spots"),
interpoly shorts (electrically conductive bridges between the polysilicon gate electrodes), etc. Eventually, the quality of the starting
material (the silicon wafers) will be the decisive factor determining the number of bad pixels, which is therefore expected to
decrease significantly in the near future with improvements in semiconductor process technology.
From this, two conclusions are drawn: Firstly, present applications of large-area CCD image sensors require suitable correction
algorithms to provide for interpolation over the large number of bad pixels and columns. Depending on the particular application,
simple repetition fill-in schemes could be appropriate, or sophisticated interpolation procedures based on information theory and
local picture statistics might be required. Secondly, the yield of large-area CCD image sensors has to be increased, for which the
semiconductor manufacturers need to make a significant effort.
Nevertheless, as a matter of curiosity, let us estimate the maximum geometry and pixel count of a CCD image sensor, fabricated
with the newest generation of semiconductor equipment, which is currently introduced into the silicon foundries. It has recently
been demonstrated, for example by the Eindhoven CCD image sensor foundry of Philips, that step-and-repeat photolithography
equipment for 6" (150 mm diameter) silicon wafers can be used for the manufacture of large, photolithographically mosaicked
image sensors. In this way a minimum feature size of 1 um or less can be employed. Using our 5x rule of thumb, this results in
pixel dimensions of about 5 um. Assuming a usable wafer diameter of 140 mm, the largest possible square image sensor measures
99 mm on a side. This would result in an image sensor with about 20,000x20,000 pixels. While the fabrication of such a large-area
CCD image sensor with 400 million pixels is in the realms of the possible today, it is doubtful that a sufficient number of working
devices without an excessive bad pixel count could be manufactured using today's state-of-the art semiconductor technology.
4. PHYSICAL LIMITATIONS OF THE PERFORMANCE OF CCD IMAGE SENSORS
The most important parameters characterizing the performance of CCD image sensors are the full well capacity Omar (De
equivalent r.m.s. readout charge noise o, the spectral quantum efficiency, the charge transfer efficiency (CTE), output sensitivity,
the maximum readout speed, the number of readout channels and the dark current. A derivative measure is the dynamic range D/R,
defined as
D/R = 20g Zana | (2)
o
The full well capacity depends on the area covered by the CCD pixels and on the applied gate voltage. For buried-channel CCDs,
the maximum charge per area is a few thousand photoelectrons per um?. This implies that smaller pixels exhibit also a smaller full
well charge capacity. The r.m.s. readout charge noise ¢ is primarily determined by the capacitance C of the charge detection node,
but also by the temperature T, the bandwidth B and the detection transistor's transconductance g ("the amplification"), according to
the following relationship [14]
| T
S&C AKT Ba (3)
8
with the Boltzmann constant k=1.38x10* [J/K], and a parameter & taking on values between 2/3 and 10, where a value close to
2/3 is more typical for practical usage in CCD output amplifiers, see Ref. [14]. This formula assumes perfect signal conditioning, to
suppress reset noise, for example by multiple correlated sampling [15]. Typical values for these parameters are C=100 fF, T=300K
(room temperature), B=20 MHz (video) and 22-1000 umho (20.001 A/V). Under ideal conditions, and assuming o=1 for
simplicity, an output noise of 12 electrons would result. The output sensitivity, often quoted for image sensors, can be computed
approximately as the charge of one electron over the transistor's input capacitance, e / C, for a source-follower type output ampli-
fier. In our example, a charge sensitivity of 1.6 WV / electron would be predicted, as observable in actual CCD image sensors.
From equation (3) it is obvious by which means the dark noise of a CCD imager can be reduced and the D/R can be increased:
1. The temperature can be decreased, i.e. the image sensor is cooled. For astronomical applications, for example, temperatures as
low as -1200C .. -809C are commonly employed. The main reason for cooling usually is that the dark current is reduced
substantially - it is halved for every reduction of the temperature by about 8K - so that very long exposure times become possible.
Dark current rates of less than one electron per hour are easily obtained in this way.
2. The bandwidth and - as a consequence - the readout rate are reduced. Unfortunately, below frequencies of about 100 kHz,
another noise mechanism comes into play (1/f instead of thermal white noise), so that in practice, slow-scan devices are operated
with optimum readout rates around 100 kHz. Obviously, much longer readout times result, that can be as long as several minutes
per frame with large-area image sensors.
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995