3. As an alternative, one could measure the same photocharge several times, non-destructively, and average over many [24 
measurements [16]. In this way, a readout charge noise of less than 0.5 electrons r.m.s. has been demonstrated. effi 
4. Could the transconductance of the output transistor be increased? No! Since this is only possible through a larger gate area, the The 
capacitance would increase correspondingly. As a consequence of this and of equation (3), the noise would actually be increased. higl 
5. The individual pixels exhibit local gain by employing a suitable photocharge amplification effect. Such devices have been tran 
demonstrated, see for example [17], but due to their complexity, they have not found wide-spread application. 979 
6. The most effective measure to reduce the readout noise is to reduce the charge detection node’s capacitance, as can easily be hig] 
seen in equation (3). This has been demonstrated for example in Reference [18], where a novel type of transistor (called a double — und 
gate MOS-FET) has been employed, showing a capacitance of less than one fF. This did not only lead to an unusually high charge CC 
conversion efficiency of 220 LV / electron, but also to a reported noise floor of only around 0.5 electrons at room temperature and Anc 
at video frequencies. MH 
In Fig. 5 the performance of this CCD image sensor is reac 
illustrated, using the values reported in [18]. Since the thar 
small pixels had a full well charge of only 4500 10 T T T T T cust 
electrons, a D/R of 79 dB results, according to equation — ,., stag 
(2). This concept has been improved by increasing the 2. obt: 
full-well charge capacity to 50,000 electrons, while 9 1l saturafion - witl 
keeping a low noise value of 1 electron r.m.s., resulting — © Alt 
in a CCD video image sensor with a D/R of 94 dB [19]. 9 ra cam 
Other schemes have been proposed, capable of reducing = o avai 
the noise and increasing the D/R, see for example [20]. g 0.1 «e 7 A le 
It is safe to predict, therefore, that the dynamic range of 3 e chai 
CCD video image sensors will be raised from today's 60 S TOO! 
dB to about 80 dB and more. If, however, the pixel size 0.01 | € 4 call 
is reduced at the same time, the full well capacity is also = alth 
decreased. In this case, shrinking the pixels is competing ‘ a tem 
with increasing the dynamic range, and - depending on ote cai) leadout noise level — $ — à 
the application - the trade-off will be different. > 
The maximum of currently achievable dynamic range equivalent number of 
can be obtained if one combines large pixels (large full noise electrons : 0.5 
well charge capacity) with noise reduction strategies 0.0001 A A A A A Wit] 
such as cooling, slow-scan output and skipper [21] 9: 1 10 100 1000 10000 100000 that 
multi-read averaging. In this way a maximum full well Number of electrons per charge packet elec 
charge of more than 500,000 photoelectrons and read-out perf 
noise of 0.6 electrons have been demonstrated [21], Trac 
representing a D/R of 6 decades, ie. 118 dB. It is a Fig. S : Ilustration of the charge/voltage conversion characteristics and larg 
demanding task to digitize this data with sufficient the extended dynamic range of a CCD image sensor with an Film 
resolution: À perfect, quantization-noise limited n-bit improved type of output detection transistor, see Reference [18] exp 
analog-to-digital converter exhibits a dynamic range of solic 
6 n « 11 [dB]. In order not to compromise the CCD's view 
D/R of 118 dB, an almost perfect 19-bit A/D converter would be required, operating at the typical lower limit of the slow-scan requ 
sampling rate of about 100 kHz as described above. pric« 
It is important to note that there is one basic noise effect CCD technology cannot affect: the intrinsic Poisson noise of light. Under the I 
constant illumination, where an average of N photons per second are detected, the individual detection event shows statistical a le 
fluctuations AN, given by the square root of N : Syste 
expe 
AN = JN (4) The 
reso| 
As an example, this means that individual measurements of 5000 photons show a standard deviation of 71 photons, already 1.4%. Ima; 
Since this noise adds to the electronic noise, it might soon dominate the total noise figure. This might prevent pixels from being 
shrunk much further. Unless, of course, "bulk" charge storage structures are employed, offering a higher full-well charge using the 
same pixel floor space. It has already been suggested that the charge storage capacity of image sensor pixels can be substantially 
increased by making use of vertically buried (“entrenched”) gate electrodes [22], similarly to the deep-reaching trench capacitors 
developed for modern DRAM technology. Since no significant research or production efforts have been reported recently, it can A ty] 
safely be assumed that such a radically different image sensor technology is not imminent. Zero: 
Other parameters of CCD image sensors are also of importance for their performance, but we will treat them only briefly in the fey 
following: 
As illustrated in Fig. 1, the spectral quantum efficiency is usually excellent, especially for wavelengths in the green and red part of Com. 
the spectrum. In some cases, however, the blue response is reduced considerably because the blue light cannot penetrate through CCD 
the CCD's electrodes efficiently enough. For these applications, it might be advantageous to illuminate the image sensor from Its pe 
behind, in the so-called back-side illuminated CCDs [23]. Alternatively, the sensor's front surface could be covered with a thin show 
sheet of scintillating phosphor material, emitting longer-wavelength (green or red) light, when near-UV or blue light is incident 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995 IAPF