Multi-function IR detector with on-chip signal processing

Transcription

1 Multi-function IR detector with on-chip signal processing Lidia Langof (1), Dan Nussinson (1), Elad Ilan (1), Shimon Elkind (1), Roman Dobromislin (1), Itzik Nevo (1), Fanny Khinich (1), Michael Labilov (1), Zipora Calahorra (1), Shay Vaserman (1), Tuvy Markovitz (1) Avi Twitto (2), Dov Oster (2) (1) SemiConductor Devices, P.O. Box 2250, Haifa 31021, Israel, (2) Israeli Ministry of Defence (IMOD), Israel, KEYWORDS: Infrared Detector, InSb, Focal Plane Array, Multi-function ROIC, on-chip signal processing, Laser Range Finder, Pulse Detection, Low Noise Imager ABSTRACT: Advanced electro-optical systems are designed towards a more compact, low power, and low cost solution with respect to traditional systems. Integration of several components or functionalities, such as thermal imager, laser designator, laser range finder (LRF), into one multi-function detector should serve this need and trend. SNIR is a new type of infrared (IR) detector, which consists of a Read-Out Integrated Circuit (ROIC) that incorporates high level of signal processing. The 0.18µm CMOS process allows for in-pixel advanced functionality, and at relatively low power consumption. The ROIC is Flip-Chip bonded to a InSb detector array with 15µm pitch. SNIR digital ROIC can be operated in either one of the following four different modes of operation: passive IR standard thermal imaging, low noise imaging (LNIM) for SWIR, two dimensional laser range finder (TLRF) in which the detector is synchronized to a laser, and asynchronous laser pulse detection (ALPD) combined with thermal imaging. We review some of the predicted and measured results for the different modes of operation. 1. INTRODUCTION Modern electro-optical systems are designed towards more compact, low power, and low cost solutions with respect to traditional systems. Integration of several components or functionalities, such as thermal imager, laser designator, laser range finder (LRF), into one multi-function detector can serve this trend. LRF is becoming an increasingly vital component to high precision targeting engagements for the imaging system user. The precise and accurate range-to-target information is an essential variable in the fire control solution of today s sophisticated weapons. Handheld military range-finders normally operate at ranges of 2 km and up to 5 km, and have about ±10m range accuracy. The more powerful models of range-finders measure distance up to 25 km and are normally installed either on a tripod or directly on a vehicle or gun platform. In the latter case, the range-finder module is integrated with thermal night vision and daytime observation system. Lasers are also used extensively as Light Detection And Ranging (LIDAR) for 3-D object recognition. With the recent advances of LIDAR technology, the accuracy potential of LIDAR data has significantly improved. State-of-the-art LIDAR systems can provide pulse repetition rate of up to 100 khz, and range measurement accuracy of few centimeters. SCD's 2-D InSb detector roadmap started in 1997 with the introduction of analog detectors with a format of 320x256 elements and 30µm pitch, and continued with large format digital detectors of 640x512 and 1280x1024 with pixel sizes down to 15µm. The next step was the development of SNIR, an 640x512/15µm InSb detector, with a Read Out Integrated Circuit (ROIC) that incorporates higher level of signal processing. The 0.18µm CMOS process allows in-pixel high functionality, and at relatively low level of power consumption. Previously [1] we presented the results of the first SNIR FPAs characterized in a laboratory Dewer. In this paper we add some of the latest results we have measured at detector level.

2 2. ROIC AND FPA The SNIR Focal Plane Array (FPA) detector is based on a 640x512, 15µm pitch, InSb array. The InSb diodes are produced in the standard and well established process of SCD's InSb diodes [2] (planar technology). The SNIR ROIC is part of the 15µm-pitch family of SCD digital output ROICs, including Pelican-D [3] and Hercules [4]. Therefore, SNIR shares peripheral circuits (outside the pixel matrix) such as readout channels, I/O circuits and controls, with its 15 µm predecessors. By controlling switches and bias levels inside the pixel, different functions can be implemented according to the predefined modes of operation. The new digital ROIC has the following modes of operation: 1. Standard IR Imaging (SIM) 2. Low Noise (very high gain) Imaging (LNIM) 3. Asynchronous Laser Pulse Detection (ALPD) 4. Two dimensional Laser Range Finder (TLRF) The Standard Imaging Mode is similar to the operation of other SCD's cooled-detectors [5], and is very similar to the Pelican-D detector [3]. The pixel signal is read via a direct injection (DI) readout circuit to an internal capacitor. Two readout capacitors are available in this mode and are selectable via serial communication. The large capacitor may be read in both Integrate Then Read (ITR) and Integrate While Read (IWR) modes, while the smaller capacitor is only operative in the ITR mode. The LNIM mode is a low noise, high gain mode designed for low flux scenarios. In this mode the diode current feeds a Capacitive Trans-Impedance Amplifier (CTIA) stage. The CTIA stage enables the use of extremely small capacitor, and provides a stabilized diode bias. For low flux application, an extremely low dark current is required. Hence, an improved low dark current InSb fabrication process is implemented, and the detector is operated at 68K. This mode may also be used for high flux and extremely short integration times (e.g. synchronous laser illumination scenarios in active imaging). In this case the low dark current is not required. Figure 1 presents the schematics of the LNIM mode pixel design. - + reset integrate Pixel Circuit x1 pixel select Figure 1. Schematic configuration of LNIM mode column wire The ALPD mode is designed to detect asynchronous short laser pulses, while providing an IR image simultaneously. An internal circuit detects short bursts of incoming flux. For each pixel, a designated bit indicates whether a burst was detected since the previous readout. The internal circuit is designed to provide an improved sensitivity as compared to detection of a laser spot by image processing. This is due to the fact that the detection circuit is hardly affected by the signal noise. The diode signal in the ALPD mode is read via a Buffered Direct Injection (BDI) circuit to an internal capacitor. Also in this mode, two different integration capacitors are available. The IWR mode is available only for the large capacitor, while the ITR mode is available for both capacitors. The BDI circuit enables improved diode bias stabilization compared to the DI readout, which is necessary at the sharp flux changes induced by the laser pulse. However, the BDI implementation comes at a cost of reducing the dynamic range (for each capacitor) by almost a third with respect to the SIM mode, and increased power consumption. Figure 2 shows the schematics of the ALPD mode pixel design. Figure 2. Schematic configuration of the ALPD mode

3 The input circuit detects current changes in realtime. An event (e.g. laser pulse) is detected by measuring a current change indicative of a radiance change during each frame. The change in current is detected by measuring the current derivative instead of integrating the current, by inputting the current into a band pass filter (BPF). The BPF removes the DC component of the input signal of the photodiode. Accordingly, the AC component is detected, thus increasing the sensitivity of the event detection. The signal is then analyzed by the signal analyzer unit including a comparator and a flip-flop circuit. The circuit of Figure 2 represents the single pixel readout within an array of pixels. The event/pulse detection is carried out simultaneously with the regular imaging mode. For example, each pixel outputs 15-bit scene data for the regular imaging and 1-bit for the pulse detection. In the TLRF mode no thermal image is created. A dedicated circuit in each pixel measures the elapsed time between the "laser-fired" trigger (which is input to the detector), and the instance in which the reflected laser beam was detected by the pixel. Thus a 2D range map of the view is created. In this mode the power consumption is relatively high. Therefore, an operation at a subwindow was implemented. Normally, the laser spot of a laser range finder covers only up to a few tens of pixels in the image. Hence operating this mode in a sub-window of 32x32 pixels is normally sufficient. The number of such sub-windows and their location in the array can be changed via serial communication. Figure 3 shows the TLRF mode design schematically. The TLRF mode circuit includes the detection mode circuit of Figure 2 (ALPD mode) for detecting the event, and additional circuit for locating the pulse event in the time scale. The additional circuit comprises the TLRF ramp, switch and readout capacitor. The TLRF ramp voltage is initiated by an external triggering mechanism and the slope is defined by the system. The TLRF ramp slope defines the time span of the measurement, which is translated to the depth or range in the 3D image. The input signal which contains the pulse causes the comparator signal to reach the threshold voltage defined by the control system. The trigger mechanism is constituted by a switch which becomes open in response to the thresholded output of the comparator (i.e. as a result of identification of the event signal). Therefore, the readout capacitor voltage becomes proportional to the time of flight. This time of flight measurement is used by the readout utility to determine a distance to the event using a ramp conversion curve. It is possible to switch between different modes (TLRF,LNIM, and ALPD) on a frame to frame basis. This allows a design of flexible operating sequence that meets the application requirements. switch Figure 3. TLRF mode schematics

4 3.1 SIM 3. DETECTOR PERFORMANCE Radiometric performance of SNIR's standard mode where presented previously [1]. We showed that in spite of the complicated SNIR pixel design, high radiometric performance is preserved. Low NETD of about 24mK is reached as in standard SCD InSb detectors such as Pelican-D [3]. In addition, Figure 4 shows that the histogram of the NETD is symmetric around the mean value and has relatively narrow distribution imaging with very short integration time. Integration time in LNIM mode can be lower than 10 µsec. The dark current should be low compared to the photocurrent. Very low dark currents of 35 fa per pixel were reached by operating the detector at a temperature of 68K. We present here the radiometric performance of the passive (long integration time) LNIM mode. All measurements were performed using F/4 aperture and a µm cold filter, while facing 1.5 µm uniform illumination at the output of an Integrating Sphere. For best performance we use pixel Correlated Double Sampling (CDS). In this spectral range the Noise Equivalent Power (NEP) is the key parameter which is used to define the sensitivity instead of NETD. Typical NEP histogram of the array is shown in Figure 5 and its 2D distribution is presented in Figure Number of pixels Mean NETD = 24mK counts NETD [mk] Figure 4. NETD histogram in Standard Imaging Mode at 50% well-fill The Residual Non-Uniformity (RNU) after 2-point correction for 20-80% well-fill capacity measured at different blackbody temperatures and constant integration time is less than 0.03% STD/full range. The pixel operability in SNIR Standard Imaging Mode was measured using the same defect identification criteria as in our standard InSb detectors. Such criteria include shorted and disconnected pixels, NETD defects, RNU defects, and other defected pixels which are not operating properly. Typical operability is better than 99.9% NEP [fw] Figure 5. LNIM typical NEP histogram.the integration time is 2 msec 3.2 LNIM LNIM mode is a low noise, high gain mode designed for low light level imaging such as SWIR imaging at night. It can be used only if the MWIR radiation is blocked by the cold filter, or in active

5 Figure 6. NEP 2D distribution map of LNIM mode. The scale is in fwatt and the measurement conditions are as in Figure 5. In Figure 7 we show the RNU of the LNIM after linear fit correction, as a function of well fill capacity measured at different illumination levels and constant integration time. Each point is an average of 64 consecutive frames. The RNU is presented in units of STD over full dynamic range. The RNU in the LNIM mode is less than 0.3% STD/full range for signals between 20-60% well-fill capacities. In Figure 8 we present the corrected image in LNIM, which shows "white spatial noise". Figure 8. Corrected image at 55% well fill for LNIM mode. The scale is in digital levels Finally, in Figure 9 we present an image captured by SNIR in LNIM mode. Irradiance level in the dark room was about 9.5x10 5 photons/sec per pixel. The spectral range in this case is µm RNU [% of DR] Figure 9. Image captured by SNIR in LNIM mode Well Fill [%] Figure 7. RNU of LNIM mode after linear fit nonuniformity correction by linear fit to the measured data 3.3 ALPD As we already mentioned, in the ALPD mode of operation the detector produces simultaneously an IR image (up to 15 bits) and the laser detection information (one additional bit) at every pixel. The IR image performance is very similar to the Standard Imaging Mode performance and will not be discussed further. In the ALPD mode we define new parameters which are important for the ALPD characterization: False Alarm Rate (FAR) per pixel

6 as the percentage of frames indicating false laser pulse detection for specific pixel (defined as Pixel FAR in previous work [1]), mean array FAR as an average FAR of all non-defective pixels, and Detectivity as the percentage of frames indicating true laser detection for specific pixel. There are two kinds of defects defined in the ALPD mode: FAR defects - pixels with FAR above certain threshold value (e.g. 20%), and Detectivity defect - pixels with Detectivity which is less than certain threshold value (e.g. 80%). We also define Operability as the percent of pixels, out of all matrix pixels, that have Detectivity greater than 80% and are not FAR defects. FAR and Operability can be tuned by the detector sensitivity threshold level, Vt. It enables increase in detector Detectivity in applications which can tolerate higher FAR level. Alternatively it is possible to change the detector Detectivity according to the scene. Figure 10 shows that the number of FAR defects decreases exponentially with increase in threshold level Vt. FAR defects [%] Threshold level [mv] Operability [%] frame size=640 x 512 frame size=128 x Laser Power [e/pixel/frame] Figure 11. Operability measured as a function of laser power for full frame and 128x128 pixels frame For most applications, which are using multiple laser pulses, it should be easy to distinguish at the system level between FAR and laser event, even with relatively high FAR level, due to the spatial and temporal distribution of the FAR defects which can be characterized in advance. No clusters of FAR defects were found at any threshold level Vt, so laser spots that cover more than one pixel should be identified easily. Moreover, a precalibrated FAR table can be stored in memory, so a change in alarm rate should indicate a real laser event. Alternatively, FAR defects can be defined and marked, such that the FAR level can be decreased dramatically. Figure 12 shows the FAR and Detectivity improvement after defect removal and following 2x2 spatial binning algorithm. The laser power in this case is ~ 800 e/pixel. Figure 10. FAR defects vs. threshold level. Optimum threshold level can be selected for each value of the laser power in order to get maximum operability. Figure 11 shows the Operability measured as a function of laser power in electrons per pixel for different frame sizes. The difference in Operability is due to weak internal power supplies that can be redesigned in future. The sensitivity of the pixel to light pulse (photon/pixel) can be calculated using the fact that the external quantum efficiency (QE) is ~ 0.8. The mean array FAR is less than 0.5% in this case.

7 Raw DATA Operability [%] Idc=84pA Idc=54pA Idc=17pA FAR defects removed 20 Idc=1pA Laser Power [e/pixel/frame] Figure 13. Operability vs. laser power for different background DC radiation levels for 128x128 pixels frame size As a final demonstration of the ALPD mode, Figure 14 shows an image captured by the first SNIR prototype camera. The 16th bit with the laser pulse detection information (in red) is overlapped in this figure with the standard image signal. 2x2 binning Figure 12. FAR and Detectivity improvement in laser spot detection by ALPD after defect removal and 2x2 binning. Since the ALPD mode pulse detection circuit is completely independent on the integration circuit, laser pulse detection does not depend on the signal level or the integration time and can operate with very high duty cycle of the frame time. On the first prototype we measured 90% duty cycle at 30Hz. Figure 13 shows Operability vs. laser power for different DC background illumination levels. The difference is negligible. Figure 14. ALPD mode image captured by first SNIR prototype. The red spot indicates laser pulse detection from the car shield. 3.4 TLRF A widespread method, known as Laser Range Finder (LRF), to determine the distance to an object is by firing a laser pulse to the target and measuring the time duration, t, between transmission of the pulse and detection of the reflected signal. In this case the detector has to be synchronized to the laser. In SNIR this is done by

8 System to SNIR triggering pulse generated by the system, which marks the start of the timer. It is also possible to operate this mode the other way around, where the detector is the Master and the laser is the Slave. The measurement setup for the experiment reported below includes pulsed laser, pulse generator, laser attenuator, beam splitter, fast diode for laser pulse monitoring, and imaging optics. The pulse generator is triggered by Frame Start pulse coming from the detector. It generates the signal at a given delay and with controlled pulse width. This signal is used to drive the µm diode pulsed laser. The laser signal was characterized by the LNIM active mode at every pixel, and the amplitude was controlled by the attenuator. Time of flight was measured by varying the laser pulse delay to imitate a real distance. The time span was varied in discrete steps between 3 to 120 µsec. In the TLRF mode, FAR, Detectivity, Operability, and FAR and Detectivity defects are defined in the same manner as in the ALPD mode. Most of the pixels reach 100% Detectivity for high enough laser pulse energy. Therefore we define another parameter, TLRF Sensitivity, as the laser power at 50% pixel Detectivity. We define also the mean array TLRF Sensitivity as the median of all pixels. The same threshold level, Vt, as in the ALPD mode, controls the FAR and Sensitivity in the TLRF mode. Figure 15 shows the mean array FAR and Sensitivity, measured for small time span of 3 µsec, vs. the threshold level, Vt. Larger time spans have FAR lower by 10%-20% for the same Sensitivity. FAR [%] TLRF sensitivity [e/pixel] increasing Vt, FAR is decreasing exponentially and so does the Sensitivity. Figure 16 shows the Operability vs. laser power for different values of threshold level Vt. It can be seen that very high Sensitivity of 1000e/pixel/frame can be reached with lower Operability (e.g. 60%) using Vt of 73 mv. Alternatively one can choose higher operability (e.g. 95%) at lower sensitivity (e.g. Vt=127mV) Operability [%] Vt=73mV Vt=100mV Vt=127mV Vt=182mV Laser Power [e/pixel/frame] Figure 16. Operability vs. laser power for different values of threshold level In TLRF mode there are two more important parameters, the spatial RNU of the range image and the range accuracy. The TLRF signal in Digital Levels (DL) is measured as a function of laser pulse delay for different time spans. For each time span new signal to range calibration is needed. Linear fit or standard two point correction can be used, but a quadratic fit improves the results, especially close to the end points. The laser pulse delay (a) and the deviation from a linear fit (b) are presented as a function of the measured TLRF signal in Figure 17. It shows that the deviation from the linear fit is less than 3 nsec for the shortest time span of 3 µsec Threshold Level [mv] Figure 15. Mean array FAR and Sensitivity in TLRF mode. The time span is 3 µsec For low Vt the detector is very sensitive and almost all pixels are defined as False Alarms. When

9 Pulse delay [µsec] Deviation from linear fit [ µsec] a) Signal [DL] 4 x b) Signal [DL] Figure 17. (a) The laser pulse delay and (b) the deviation from linear fit as a function of TLRF signal. The time span is the shortest, 3 µsec The spatial RNU of the 2D range map was calculated for the 32x32 pixel window for different time spans. The measurement and analysis procedure is as follows. First, a set of TLRF signal measurements as an average of 100 consecutive frames is recorded for different laser pulse delays. This set was used to find the quadratic fit for each pixel. Then, the fit was applied to a second set of measurements where only one frame was recorded every time. The RNU of the range was calculated as a spatial 2D standard deviation of the corrected TLRF 2D map for every delay. The RNU is less than ~0.1% of the time span for the whole dynamic range. Figure 18 presents the corrected time 2D map of the shortest time span, within a 32x32 sub-window of the array. We find no spatial pattern in the corrected TLRF signal Figure 18. Corrected 2D map (32x32 sub-window) of measured time delay for 2.2 µsec laser delay. The scale is in µsec and the time span is 3 µsec. The detector TLRF signal was also measured as a function of the laser pulse intensity in the range of e/pixel, at the same laser pulse delay. For low laser pulse intensity below 5000 e/pixel there is an additional delay caused by the change in pixel time response. If the intensity is not known, the uncertainty in the measured time of flight can be about 0.2% of the time span. For applications with repetitive laser pulses, the detector can be operated sequentially in the active LNIM and TLRF modes, such that intensity measured by the active LNIM can be used to correct the additional delay in TLRF signal for low intensity pulses. Typically, this additional delay is proportional to the inverse of the laser pulse intensity. Tab. 1 summarizes the SNIR performance in TLRF mode. In order to achieve high range accuracy, one can use the larger time span (i.e. 120 µsec) in order to study the scene. Once a target is identified, the user can shift the operation to small time span (3 µsec), such that the range accuracy of the measurement can reach 0.3 meter.

10 Table 1. TLRF mode range performance for short, medium, and long time spans. Time Span [µsec]/ [km] RNU [nsec]/ [m] Range Accuracy [nsec]/[m] Range Accuracy for high or known laser pulse intensity [nsec]/ [m] 120µsec 18km 80nsec 12m ±150nsec ±22.5m ±60nsec ±9m 30µsec 4.5km 25nsec 3.75m ±50nsec ±7.5m ±20nsec ±3m 3µsec 450m 3nsec 0.45m ±9nsec ±1.35m ±2nsec ±0.3m We performed system level simulation to find a maximum target range that can be measured with the TLRF mode as a function of laser pulse energy. Some basic assumptions were made: 50mm optics diameter, 50% total transmission, 20% object reflectance, and 1µm laser wavelength. For high Operability we choose a threshold irradiation level of 1600e/pixel/frame. The simulation was done for different laser divergence angles, which is equivalent to the laser spot size at the FPA. Figure 19 shows the simulation results for different laser spot sizes [pixels]. Range [m] pixels 33 pixels 74 pixels 205 pixels 524 pixels 991 pixels Laser Pulse Energy [mj] Figure 19. Calculated range vs. laser pulse energy for different laser spot sizes at the FPA (in pixels) And finally, Figure 20 shows ALPD gray scale image with TLRF data in color overlapped on image. The data was measured by switching between ALPD and TLRF modes. Two laser spots are detected, one from the tree (16 meters) and one from the car (23 meters) Figure 20. ALPD gray scale image with TLRF data in color overlapped. TLRF scale is in meters. 4. SUMMARY In this paper we presented SNIR, SCD's advanced multi-function IR detector. We showed that SNIR preserve a high quality standard imaging in the MWIR together with new functionalities. It meets the challenge of implementing additional functions in the pixel on large format and small pitch ROIC. The high functionality has been achieved by using 0.18µm CMOS technology. SNIR enables the development of small and compact systems that include functions of IR imager, laser range finder, laser designator finder and low light imager. We showed that the measured range accuracy of the laser range finder mode (TLRF) is similar to the current commercial LRFs and has an additional value of IR image and LRF signal spatial correspondence and target recognition. Moreover, for short time spans 2D accurate range map can be obtained by multiple laser pulses operation. This enables 3D imaging with 30 cm depth resolution. Low noise readout mode (LNIM) which is suitable for active or passive imaging in the SWIR was also presented. Active

11 LNIM synchronized to a pulsed laser can be used with very short integration time lower than 10 µsec, and therefore can be used in parallel to other modes in the MWIR. Very low readout noise and dark current enables long integration time for the passive low light imager (SWIR only configuration). Finally, the ALPD mode can be used to acquire IR image simultaneously with see-spot function to identify pulses from laser designators or any other fast events. Unlike the TLRF mode where the pulse duration should be less then 25 nsec for maximum Sensitivity, the ALPD mode is sensitive to events duration of up to tens of µsec. 4.1 Acknowledgement The development of SNIR was funded by the Israel Ministry of Defense (IMOD). We would like to thank SCD's personal who were involved in the development: Aharon Shechter, Ofer Nesher, Yael Schlesinger, Itay Hirsh, Nir Fishler, Lior Shkedy, Miriam Geva; Anat Liran; Ernesto Kuznitzky; Avi Adato; Ofek Levin; Igor Sapiro; Ariel Cohen; Amnon Adin REFERENCES [1] Lidia Langof, Dan Nussinson, Elad Ilan, Shimon Elkind, Roman Dobromislin, Itzik Nevo, Fanny Khinich, Michael Labilov, Zipora Calahorra, Shay Vaserman, Tuvy Markovitz, Ofer Manela, David Elooz, Avi Tuito, Dov Oster, " Advanced multi-function infrared detector with on-chip processing" Proceedings of SPIE vol. 8012, Infrared Technology and Applications XXXVII Conference, April 2011, pp 80120F-80120F-13 [2] O. Nesher and P. Klipstein, "High performance IR detectors at SCD present and future", Proceedings of SPIE vol. 5957, Infrared Photoelectronics, August. 2005, pp 0S1-OS12 [3] T. Markovitz, I. Pivnik, Z. Calahorra, E. Ilan, I. Hirsh, E. Zeierman, M. Eylon, E. Kahanov, I. Kogan and N. Fishler "Digital 640x512 / 15micron InSb detector for high frame rate, high sensitivity and low power applications", Proceedings of SPIE vol. 8012, Infrared Technology and Applications XXXVII Conference, April 2011, pp 80122Y-80122Y- 10 [4] O. Nesher, I. Pivnik, E. Ilan, Z. Calahorra, A. Koifman, I Vaserman, J. Oiknine Schlesinger, R. Gazit, and I. Hirsh, " High resolution , 15 µm pitch compact InSb IR detector with on-chip ADC ", Proceedings of SPIE vol. 7298, Infrared Technology and Applications XXXV Conference, April. 2009, pp 72983k-1-9. [5] J. Oiknine Schlesinger, Z. Calahorra, E. Uri, O. Shick, T. Fishman, I. Shtrichman, E. Sinbar, V. Nahum, E. Kahanov, B. Shlomovich, S. Hasson, N. Fishler, D. Chen, T. Markovitz, " Pelican SCD's /15 µm pitch InSb detector" Proceedings of SPIE vol. 6542, Infrared Technology and Applications XXXIII Conference, July. 2007, pp