1. Introduction

An intensified charge coupled device (ICCD) camera is one of the promising types of matrix detectors with single-photon sensitivity. In general, an ICCD camera consists of a photocathode, a micro-channel plate (MCP), a phosphor screen, a CCD imaging chip accompanied by readout electronics, and an optical coupling system between the latter two. Such a design enables certain flexibility: for instance, by selecting the photocathode material, one can achieve a higher sensitivity in a preferred spectral range [1, 2]. The high voltage applied to the MCP serves for the multiplication of electrons released from the photocathode; by tuning its value, one can change the gain of the camera and this way improve the single-photon sensitivity. Moreover, the high voltage can be applied during time intervals (gates) as short as a few nanoseconds and serves therefore as a fast electronic shutter. Since the late 1970s, when the first devices were developed, the family of ICCD cameras has been widely implemented in astronomy and space research [2–5], chemistry, biology, and medicine [6]. Nowadays, ICCD cameras are more and more often used in quantum optics [7–13]. Because an incident light beam can be spread over a large number of pixels, an ICCD camera can be considered as a multi-channel photon-number resolving detector, albeit with low quantum efficiency. There are several other types of cameras with similar applications, for instance EMCCD [14–17] and intensified sCMOS cameras (I-sCMOS) [18, 19]. Compared to an EMCCD camera, an ICCD camera has significantly lower dark noise [20, 21] and therefore enables detection of single photons (‘on/off detection’). As to I-sCMOS cameras, they have the same design as the ICCD cameras but with the CCD chip replaced by a sCMOS chip.

An essential part of detecting single photons with an ICCD camera is distinguishing a single-photon event from the dark noise. As in most photon counting devices, this is done by applying a threshold for the charge read out from the output CCD. This charge has broad, and considerably overlapping probability distributions for the cases of dark noise and the incident photons at the input. For this reason, the value of the threshold affects both the signal-to-noise ratio (SNR) and the quantum efficiency (QE). For threshold values allowing a sufficiently high SNR, the QE is therefore significantly less than the one of the photocathode.

Here we measure the quantum efficiency of an ICCD camera operating in the single-photon detection regime, as a function of the threshold value. The measurement is performed using the absolute calibration method [22–26], with the radiation of spontaneous parametric down-conversion (SPDC) at the input. The absolute calibration technique is applied in an autonomous manner, i.e., with the ICCD camera used as both the reference detector and the detector under test (DUT). We also study the quantum efficiency versus the wavelength of the incident radiation as well as its uniformity over the camera matrix.

2. Single-photon detection with ICCD

In the single-photon detection mode, the ICCD camera works as a gated on-off detector. During the gate time, high voltage is applied to the MCP, and once an incident photon releases an electron from the photocathode, this electron will be multiplied in the MCP and create a light flash on the phosphor screen, which, in its turn, is sent to the output CCD with a fiber bundle. As a result, a charge is created in some pixels of the CCD chip. Due to the inevitable crosstalk, it is never a single pixel that is affected by a single photon at the input; this crosstalk can be reduced by combining m×n pixels into ‘superpixels’ (m×n binning). The CCD charge is finally read out digitally. This value is somewhat higher than the background, which mainly comes from the electronic noise of the CCD and the background luminescence of the phosphor screen. As illustrated in Fig. 1a, the charge read out from an array of pixels (blue solid curve) shows higher values for some pixels compared to the background. The latter shows fluctuations between −25 and 25 output analog-to-digit converter (ADC) values, after the average value has been subtracted. In this situation, one uses a variable threshold parameter S_th to distinguish the signal caused by an incident photon from the background noise on every pixel of the camera (Fig. 1a). Whenever the charge read out from the output CCD exceeds the threshold, it is assumed that the corresponding pixel detected a photon. It should be emphasized that this strategy is a binary processing and cannot resolve several photons detected by the same pixel within one exposure time. Hence it is necessary to reduce the input photon rate until the probability of two photons hitting a single pixel or group of pixels becomes negligible. Technically, this can be done by reducing the exposure time. Importantly, the thresholding process takes place after the data acquisition, and hence the threshold value can be chosen after the actual experiment.

 

Fig. 1 (a) The readout values of an array of pixels in the ICCD camera (blue solid curve) and the threshold parameter S_th (red dashed line). Every peak that is above the threshold is interpreted as a photon detection event. (b) The mean number of photon detections on an area formed by 9 × 9 pixels per 100ns gate with SPDC radiation (blue solid line) and no radiation (black dashed line) at the input, as functions of the threshold value. Inset: Signal-to-noise ratio (SNR) as a function of the threshold value.

Download Full Size | PDF

In order to optimize the single-photon detection ability of an ICCD camera, it is important to choose the threshold correctly. Figure 1b shows the average number of counts per exposure versus the threshold value for two cases. In the first case, the ICCD is exposed to the incident SPDC light (blue points) and in the second one, the ICCD input is blocked, and only noise is present (black points). Note that here and further on, the threshold values are shown after the subtraction of the ‘background’ value, which was obtained in the absence of illumination. This value varied randomly from pixel to pixel within the range from 600 to 650 ADC values. One can see that, as the threshold value increases, the noise reduces faster than the signal, hence the signal-to-noise ratio (SNR) grows, as shown in the inset in Fig. 1b. But above the threshold of roughly 80, SNR grows very slowly and practically reaches a constant value. This behavior can mean that the dark noise is almost completely suppressed, and the remaining noise results only from the stray light of the environment. In this threshold range, the detection events are reliable.

3. Autonomous absolute calibration of the camera

In addition to reducing the noise, an increase in the threshold value leads to the reduction of the QE. For this reason we performed an absolute calibration of an ICCD camera using the Klyshko method [22–24, 26], which is based on twin-photon correlations in SPDC. The method is absolute since it requires no knowledge of the reference detector QE for calibrating the DUT.

It should be stressed that although ICCD cameras are widely used as ‘on/off’ detectors [9–13], their QE has been never measured in this mode. Perina et al. [27] characterized the QE of an ICCD camera in the quasi-analog regime, where more than 2 photons were registered by a single pixel on the average, using a technique involving the measurement of a set of statistical moments. However, it is important to characterize the QE in the truly single-photon regime, which is required, for instance, for registering anti-bunching [13].

In the case of pulsed radiation, the QE of the DUT can be expressed as [28]

The experimental setup (Fig. 2a) uses type-I SPDC with collinear, nearly frequency degenerate, phase-matching in a 3 mm BBO crystal pumped by a CW diode laser beam (wavelength 405 nm, mean power 80 mW). The calculated wavelength-angular spectrum is shown in Fig. 2b. One can see that, although the maximum intensity should be observed in the collinear direction, the spectrum has a large angular width at each wavelength, which enables spatially resolved measurements. Because of its broadband background spectrum, the pump is additionally filtered by a bandpass filter (BF1, central wavelength 405 nm, FWHM 10 nm, and transmission 95%). A long-pass filter (LF1, cut-off wavelength 450 nm, transmission 95%) after the BBO crystal reflects the pump beam back. The SPDC radiation is far-field projected onto an ICCD camera (PI-MAX3, Princeton Instruments) with the help of a lens (transmission 99.7%) with 250 mm focal length. After the lens, another long-pass filter (LF2, cut-off wavelength 650 nm, transmission 95%) is used to further suppress the SPDC radiation of irrelevant wavelengths.

 

Fig. 2 (a) Experimental setup: a CW diode laser beam is filtered by a bandpass filter BF1 and pumps SPDC in a BBO crystal; combination of a polarizing beam-splitter (PBS) and a half-wave plate (HWP) is used to measure the noise caused by fluorescence and stray light. A long-pass filter LF1 suppresses the pump beam after the crystal; the SPDC radiation is far-field imaged onto the ICCD camera with a lens. After the lens, a long-pass filter LF2 filters out irrelevant wavelengths, and two bandpass filters (BF2 and BF3) cover different halves of the camera field of view for choosing the ‘reference’ and the ‘DUT’ wavelength bands. (b) Calculated wavelength-angular spectrum for the collinear slightly non-degenerate SPDC. (c) A typical image taken by ICCD by acquiring 4.8 million frames with the gate time 100 ns and the threshold 70, with the ‘reference’ filter BF2 selecting a bandwidth of 10nm around 780nm and the ‘DUT’ filter BF3, a bandwidth of 40nm around 850nm. The pixels inside the red square are chosen as the ‘reference’ detector. (d) The result of the g⁽²⁾ measurement between the ‘reference’ detector and all the other pixels of the camera, the threshold being 80. The filled red square marks the reference detector (the same as in panel c) and the empty green square shows the chosen ‘DUT’.

Download Full Size | PDF

For achieving the highest photon sensitivity, the MCP gain of ICCD camera is set to the maximal value of 100. The ICCD camera is gated by switching on the MCP high voltage. Therefore, even though we use a CW pump, the photon count number per frame, which is measured by the chosen detectors (groups of pixels) is fully equivalent to the photon count number per pulse, which is required in Eq. (1). Similarly, the gate time is equivalent to the coincidence window [29, 30].

An important issue is the choice of the binning (superpixel size) and the gate time. With too many pixels combined into superpixels, the spatial resolution is too low. On the other hand, because the rate of acquiring frames is restricted by the readout speed, which, in its turn, depends on the number of superpixels in the frame, the binning influences the frame acquisition rate. With superpixels too small, the acquisition rate is too low. In our experiment, the binning is 8 × 8 pixels, which leads to an acquisition rate of 26Hz. Because in the single-photon counting mode one has to reduce the probability of two photons hitting the detection area, the average number of photon detections per frame should be much less than unity. To provide enough statistics, we therefore acquire more than three million frames for each experiment, which leads to more than 20 hours of data acquisition.

The gate time is chosen in such a way that the number of counts per detection area is between 10⁻⁴ and 3·10⁻¹, depending on the threshold value. Besides, with the gate time too large the ratio of the number of total coincidences N_cc to the number of accidental ones N_acc (coincidence-to-accidental ratio, CAR) is too low due to the increase in the number of detected temporal modes. This leads to an increase in the uncertainty of measurement using Eq. (1). The value of CAR, equal to the second-order normalized intensity correlation function g⁽²⁾, can be directly inferred through the measurement of the latter. For 100 ns gate time, is as high as 20 (Fig. 2d).

In order to calibrate the ICCD camera in an autonomous regime, we use its different groups of pixels as the reference detector and DUT. In the setup, we place two different filters in front of the input window of the ICCD camera, one (BF2) covering the left-hand half of the camera and the other one (BF3) covering the right-hand half. A typical image accumulated over 4.8 million frames with the gate time 100 ns and the threshold 70 is shown in Fig. 2c. In the setup the BF2 filter is centered at 780 nm (FWHM 10 nm, transmission 94%) and the BF3 filter is centered at 850 nm (FWHM 40 nm, transmission 98%). In this situation, the left-hand side pixels can only detect the ‘reference’ photons within the 10 nm band around the chosen wavelength, and the right-hand side pixels can detect photons from a broader wavelength range, involving the matching wavelengths. In each experiment, we choose a group of pixels (red square in Fig. 2c) as the ‘reference’ detector. To demonstrate SPDC correlations, we have measured the value of the normalized second-order intensity correlation function g⁽²⁾ $(\overrightarrow{r})$, $\overrightarrow{r}$ being the position vector, for the chosen reference detector and all the other pixels of the camera. The result, shown in Fig. 2d for the threshold S_th = 80 and the gate time 100 ns, indicates the area where the correlated photons could be found. The corresponding values of g⁽²⁾ are as high as 20, showing a good CAR. The high values of g⁽²⁾ around the reference detector (marked by a filled red square) are due to the cross-talk, and the spikes in the central area of the image are within the uncertainties, which are high due to the low illumination level in this area.

Because in the absolute calibration method all photons detected by the reference detector must be accessible to the DUT, the latter (green square in Fig. 2d) is chosen larger than the area showing photon correlations. Also, the bandwidth of the ‘DUT’ filter is broader than the one of the ‘reference’ one (4 times broader in the case shown in Fig. 2c). The bandpass filter in the DUT channel is used for preventing high count rate in the DUT, which occurs in this channel even with the relatively low photon flux and short gate times. This is caused by the large size of the DUT, which, without additional spectral filtering, would cover far larger spectral range than necessary.

It is important that for the calibration, we choose collinear non-degenerate phase-matching (Fig. 2b). The reason is that in this case, a certain solid angle selected by each of the DUT pixels corresponds to a narrower wavelength range than in the case of a non-collinear phase matching. As a result, the saturation of the DUT is reduced.

Another point is the measurement of the number of accidental coincidences N_acc and the number of noise counts N_n in Eq. (1). For finding N_acc, one should measure the number of coincidences between either spatially displaced areas or frames delayed in time. The first option is not convenient due to the spatial inhomogeneity of the SPDC radiation. Therefore, we take as N_acc the number of coincidences between the counts of the DUT and the reference detector acquired in different frames, separated by tens of milliseconds or more, and therefore have no correlations. For measuring the number of noise counts, we use in the pump beam a polarizing beam-splitter (PBS) followed by a half-wave plate (HWP). Normally the pump has extraordinary polarization in the BBO crystal. When we rotate its polarization by 90°, the phase-matching condition is no longer satisfied and hence no SPDC light is generated, but the level of fluorescence and stray light is the same. Therefore, the number of detected events for the 45° orientation of the HWP is used as N_n in Eq. (1). It includes all sources of noise: dark noise, fluorescence, and ambient light [31].

It should be noted that the absolute calibration method yields the quantum efficiency of the whole optical channel, including the transmissions of all elements from the light source to the detector [24], which is estimated as 88 ± 2% in total. All the results given further in this paper have been corrected for this 12% loss in the optical channel.

Finally, in Table 1 we present the various sources of uncertainty in this measurement. The uncertainties are primarily statistical, due to the low frame rate. For this reason, we do not consider here sources of much less significance (like in Ref. [25] where record low uncertainties about 0.15% were reached). In our experiment, the main source of uncertainty is the limited number of coincidence events. As expected, the statistical distribution of both single-photon counts and coincidences in our experiment follows the Poissonian statistics, which leads to the accuracy of 20% (one standard deviation) if no additional averaging is performed. The uncertainty is reduced to 5% if we average additionally over several groups of pixels. Table 1 presents the data for two different threshold values, since the threshold is an important parameter of the experiment. Listed in the table are typical mean values and uncertainties of all quantities entering Eq. (1). The data on N_c, N_acc, and N_Ref are acquired over at least 3.5 million frames and the data on N_n, on about 70 thousand frames. (This number is sufficient due to the low number of noise counts and the small corresponding uncertainty ΔN_n.)

Table 1. Typical MeanValues and Uncertainties of Quantities used in Eq. (1), as well as of the Transmission T of Optical Elements in DUT Channel

View Table

4. Results and discussion

Figure 3a shows the QE (blue triangles) measured for a range of threshold values from 45 to 120 for the wavelength 790 nm ± 5 nm, which is conjugated to the wavelength 830 nm of the chosen reference filter (FWHM 10 nm, transmission > 95%), and filtered by a broadband filter (central wavelength 800 nm, FWHM 40 nm, transmission 99%) on the DUT side. As one would expect, the measured QE decreases as the threshold increases. At threshold values smaller than 45, the measurement error is too large because the incident photons cannot be distinguished against the strong background noise (see Fig. 1b).

 

Fig. 3 (a) The QE measured at the wavelength 790nm ± 5nm as a function of the threshold value. Blue triangles are measured through absolute calibration. For reducing the measurement uncertainty, averaging over 43 groups of pixels is performed. Black circles are obtained by measuring the mean number of detected photons for a single DUT pixel, with the noise subtracted, as a function of the threshold value, and assuming that the QE value coincides with the one measured at threshold 100 (shown by red dashed line in panel a) by the absolute method. (b) Measured QE for the threshold value 100 as a function of the wavelength (red triangles). The blue circles show the QE of the photocathode given by the manual. The two datasets show a similar tendency. (c) QE measured for 43 different pixels for the same threshold 60 at the wavelength 790nm ± 5nm. (d) The values of QE (with the error bars) plotted versus the vertical coordinates of the pixels with approximately the same horizontal positions (shown in panel c by red outline).

Download Full Size | PDF

It is worth noting that, as long as the QE is measured in an absolute way for one threshold value, its measurement for the other threshold values can be performed in a much easier way, as a relative measurement. For this, the mean number of photon count events per single pixel within the DUT, after subtracting the background noise, for the same acquired dataset, is measured as a function of the threshold value (black circles in Fig. 3a). The QE resulting from this relative measurement is rescaled in such a way that at high threshold values (we chose a value of 100), where the SNR reaches a constant value, it agrees with the result of absolute calibration. Comparing the two curves in Fig. 3a, we see that the absolute measurement underestimates the QE at lower threshold values. This can be explained by the high count rate of the DUT, not only due to the incident light, but also due to the the electronic noise. This effect is similar to the dead-time effect, or saturation of on/off detectors, and it certainly reduces the measured QE. In experiments where a smaller number of pixels are used as the single-photon detector this problem of saturation is less crucial. Because the relative measurement (black points) is taken with just one pixel, the ‘saturation’ is less pronounced and the QE is higher at small threshold values.

The maximal QE value measured this way is 8.5% ± 0.5% at a threshold of 50. This value is considerably lower than the one given for the photocathode of the intensifier (Unigen II filmless Gen III, Princeton Instruments) for the same wavelength, namely 28% [32]. This is in full agreement with the fact that, as stated above, the necessity of thresholding reduces sufficiently the QE with respect to the value provided by the photocathode. However, because the thresholding reduces the QE regardless of the incident light wavelength, the spectral dependence of the QE is only determined by the photocathode sensitivity.

As an additional test, we performed a measurement of the QE with an independently calibrated source. As such, we used the radiation of a diode laser at the wevelength 640 nm, attenuated to 4.32 ± 0.02 μW. After further attenuation with a calibrated filter (an ND filter, nominally OD6), we reduced the photon flux to 1.4 ± 0.2 photons per 100 ns frame, distributed over approximately 300 pixels. A measurement of the beam profile on the camera allowed us to infer the mean number of photons hitting a given pixel. From this, the QE was found to be 5 ± 0.2% at threshold 100. This is somewhat larger than the absolute measurement gives for 790 nm (4.1 ± 0.2%), in full agreement with the QE wavelength dependence.

The dependence of the QE on the wavelength was measured with the same setup, by changing different reference narrowband filters. In total, we used 4 filters, each one with 10 nm FWHM bandwidth, centered at 770 nm, 780 nm, 810 nm, and 830 nm. In each case, the corresponding DUT wavelengths were covered by one of the two bandpass filters: 800 nm filter (FWHM 40 nm, transmission 99%) and 850 nm filter (FWHM 40 nm, transmission 98%). The obtained values of QE are plotted in Fig. 3b (red triangles) for the threshold S_th = 100. The wavelength dependence of the QE is in a good agreement with the photocathode QE dependence (blue circles in Fig. 3b) [32]. Clearly, the QE values measured with this relatively high threshold are much lower than the photocathode QE. This is caused by the fact that the overall QE reduces considerably with the threshold value (Fig. 3a), and is therefore much less than expected from the properties of the intensifier alone. At the same time, the measured wavelength dependence of the QE repeats the one of the photocathode.

Finally, Fig. 3c shows the distribution of the QE over different pixels for the threshold 60 at the wavelength 790nm ± 5nm. The plot shows some non-homogeneity in the vertical direction; however it is almost within the measurement uncertainty. This can be seen in Fig. 3d where the QE is plotted versus the vertical coordinate for pixels with approximately the same horizontal coordinates (shown in panel c by red outline). We do see a slight decrease of the QE from top to bottom (left to right in panel d) but it is within the measured 20% uncertainty. This allowed us to average the QE value plotted in Fig. 3a over all regions shown in Fig. 3c and thus reduce its uncertainty to about 4% at S_th = 60.

5. Conclusion

In conclusion, using the absolute calibration method, we have measured the quantum efficiency of an ICCD camera. Although such cameras are now often used for single-photon detection in quantum optics experiments, no absolute measurement of quantum efficiency using the Klyshko method has been performed so far. Such a measurement is especially important because a single-photon detection event in an ICCD is distinguished from the noise by setting a threshold, and the value of the threshold has a strong effect on the quantum efficiency. We have performed the absolute calibration of the camera in an autonomous regime, so that different parts of the camera were used as the detector under test and the reference detector. The QE has been measured as a function of the threshold value and the wavelengths. Because at low threshold values, application of the method leads to the saturation of the group of pixels used as the detector under test, we also made a relative measurement of the quantum efficiency. The maximal value observed from the absolute measurement for the wavelength 790nm ± 5nm was around 7.0% ± 1.5% at low threshold while the relative measurement resulted in 8.5% ± 0.5% in this case. In addition, the inhomogeneity of the camera was assessed by measuring the QE of different pixels. Our results have shown that with any reasonable threshold, the overall single-photon detection efficiency is about three times smaller than the QE of the photocathode. At the same time, the wavelength dependence of the measured QE repeats the spectral sensitivity of the photocathode.