Serguei Palto
Senior Member
- Messages
- 1,378
- Solutions
- 1
- Reaction score
- 1,441
Recently, there has been a lot of discussion regarding the noise performance of the G9M2. In this post, I want to share my thoughts on the G9M2, which are based on the results of noise measurements of the G9M2 in various modes, as well as on comparison with other cameras.
This work focuses on the dynamic range of four cameras: G9M2, G9, OM-1M2 and OM-1.
It should be noted that these results were made possible thanks to the tremendous work done by JRSFORUMS, which provided data for the G9M2 not only in the form of black-frames, but also white-frame data for measurements of linearity response on the basic ISO, gain and photon noise for various ISOs. Black-frame data for measuring the dynamic range of the OM-1M2 was kindly provided by LOKATZ. I express my deep gratitude to these people.
DISCUSSION
Let's start by showing the dynamic range measurements of the G9M2 in two modes: i) DR boost mode; ii) SH20 (non-boost) mode, Plot #1.
Plot #1. G9M2 dynamic range in boosted and SH20 (non-boosted) mode.
As one can see, in the boost mode at low ISO values, the DR is more than 2 EV higher than in SH20 mode for all three RGB channels. The boost effect only slightly pronounced above ISO 3200, but small boosting advantage of about 0.3EV remains even at high ISO values up to ISO 25600.
Plot #2 shows the measurements for the green channel of the G9M2 compared to the G9.
Plot #2. DR data for G9M2 in comparison with G9. Both are with ES at 1/32000 s.
As you can see, the dynamic range of the G9M2 at the base ISO 100 reaches a value of 14 EV. This value was tested for three different G9M2 cameras that were available to JRSFORUMS, under different conditions of obtaining black-frames. The reason for this check was that for the black-frames provided earlier by Christian Comes, the measured DR for the green channel is over 15 EV. Perhaps the reason for the differences in the maximum DR values at the base ISO between the current and earlier measurements is due to the changes made by Panasonic in the updated firmware version of the G9M2 camera.
The GM2 value of 14 EV is one stop higher than the value of 13 EV measured for the G9 at the base ISO 200. The dynamic range of the G9M2 exceeds that of the G9, and as you can see from the Plot #2, this excess is at least 0.5EV throughout the all-ISO settings.
It should be noted that the ISO 100-160 range for the G9 is declared by Panasonic as extended, although our measurements show that the analog gain at these ISOs is decreased compared to the gain at the base ISO 200. For example, the analogue G9 gain at ISO 100 is half that of ISO 200. It is also important that the maximum response values at ISO 100 reach the maximum ADC values minus the baseline offset. The only feature of ISO 100, in comparison with ISO 200, is a certain nonlinearity of the response with increasing exposure, which becomes noticeable as the response increases and approaches the maximum possible value. Thus, ISO 100 in G9 can be regarded as the true ISO if the nonlinearity of the response is neglected.
For the G9 at ISO 16000, 20000 and 25600, the measurements showed no change in the analog gain compared to its values at ISO 8000, 10000, 12800, respectively. In Exif, these ISO values (16000 – 25000) are given a double digital gain. Since digital gain always results in a proportional increase in read noise, this fact was taken into account in the data shown in #2 for the G9.
Depending on the ISO, the data for the G9 show the "oscillating" behavior of the measured DR. Since the measurements showed a linear increase in gain with an increase in ISO, this behavior can be explained by the peculiarities of the electronic circuit of the analog amplifier with programmable gains. At certain intermediate gains between the “principal” 200, 400, 800, 1600, 3200, the electric circuit has improved noise characteristics. We will see a similar, "oscillating” behavior below for the OM-1M2 camera as well.
For the G9M2, when switching from ISO 3200 to ISO 4000 (Plot #2), one can see that there is no change in DR. Many sources claim that since ISO3200, Panasonic has applied noise reduction. Our “white-frames” measurements shown in Plot #3 do not support this view. Indeed, at the transition from ISO-3200 to ISO 4000, the read noise does not change, despite the fact that the gain increases in proportion to the growth of the ISO value, and this can be interpreted as some (0.3 EV) noise reduction. However, if we look at the dependence of the total noise (read plus photon noise), where the main contribution is related to the photon noise, we can see that when moving from ISO 3200 to ISO 4000, the noise increases from a value of 117.5 to a value of 148.5. That is, the noise increases by 148.5/117.5 = 1.264 times. At the same time, from data #2, the gain increases from 31.5 to 40.0, that is, 40/31.5 = 1.27 times, which is in very good agreement with the increase in photon noise. Thus, there is no noise reduction for ISO over 3200, and the observed feature should be attributed either to the lower read noise which is due to changing characteristics of the electronic circuit, as in the case of G9, or the features of dynamic range boosting. Similar behavior, when the circuit read-noise performance depends on the gain, can be seen in the example of the OM-1M2 in the ISO ranges 200-800, 800-1000, 10000-12800, see Plot #4 and #5.
Plot #3. Dependence of the ISO100-normalized gain (black line plus symbols), total (red scatters) and read (green scatters) noise vs ISO settings at fixed shutter speed and lens F-number.
Plot #4. DR data for G9M2 in comparison with OM-1M2.
From Plot #4, you can see that the G9M2 has a significant advantage over the OM-1M2 in the ISO range of 100 to 800, which is about one stop (EV). However, for ISO above 12800, there is only very slight difference of about 0.15 EV between the G9M2 and the OM-1, which is comparable with the measurement error estimated as 0.07EV.
It is worth to mention that the OM-1M2 DR is found to be different for the case of electronic and mechanical shutter, Plot #4. Also, it was found that in case of the mechanical shutter the DR for OM-1M2 is very similar to DR for OM-1, as one can see in Plot #5. By the way, it is easy to see from the presented data that at ISO switching in the intervals of 250 - 320, 800-1000, 12800-16000 there are pronounced jumps (growth) of DR, which could be interpreted as the noise reduction by analogy with G9M2 in the range of 3200-4000. However, of course, such an interpretation would be flawed, as shown by the increase in photon noise in case of G9M2.
The high dynamic range mode implemented in the G9M2 does not exist in all regimes, but depends on the low shutter speed and ISO, as shown in Plot #6.
As you can see, the G9M2's boost Dynamic Range mode is implemented throughout the ISO range only for shutter speeds exceeding 1/15 sec. If the ISO does not exceed 500, the corresponding shutter speed is greater than 1 reverse second for the boost DR mode.
Plot #5. DR OM-1M2 vs OM-1 (green channel; mechanical shutter).
Plot #6. Transition (red arrows) to the non-boost DR mode at low shutter speed for G9M2.
Plot #7. The G9M2 boosted DR mode at shutter speed 1/32000s vs the non-boosted modes at an SS of 2 sec and 1/32000 s for the SH20 burst mode.
The boost DR mode is off in SH20 regime, as can be seen in Plot #7. In the low ISO (100 -500) region, the difference in dynamic range between the Enhanced DR mode and the SH20 is greater than 2 EV. In the SH20 mode, the use of two different analog amplification circuits is well pronounced. It can be seen that when switching ISO from 500 to 640, there is a sharp increase in DR by 1 EV, which is explained by the use of the high-gain (G2) electrical circuit with the improved noise performance of the analog amplifier. The low-gain circuit (G1) used in the ISO range 100 -500 allows for true ISO 100, but it is characterized worse read-noise performance compared to the G2 circuit. At low shutter speed (2 s) the non-boost DR performance is different from that in SH20 mode, which again can be associated with some differences in electric circuit, which can be related to different pixel response time necessary in the both cases.
According to SH20-data in Plot #7 the DR at ISO 3200 is just equal to the boost-mode DR. From Plot #6 one can also see that at ISO3200 the DR-jump at 1/15 s is actually absent withing the measurements error. Thus, the noise performance of the G2 circuit at ISO 3200 plays a special role in the DR boosting. Indeed, the data for input-referred noise in Plot #8 clearly demonstrate that the ISO3200 is a “sweet” point at which the input-referred noise is minimal and equal to 0.121. With such a low noise the base-ISO DR in theory can be boosted up to 16 stops by digital dividing the G2 ADC data by a gain of 32. In practice with 12-bit ADC, if the G1 ADC data occupy the highest binary digits of the 16-bit register we have only 4 lowest bits free to work with for the boosting, so the G2-data can be easily divided only by a gain of 16 or lower, so the boosted base ISO DR will be lower than 16 stops.
Plot 8. Input-referred noise measured for the non-boost SH20 mode.
Plot #9. DR for G9M2 in comparison with G9 for R and B channels.
Plot #10. DR for G9M2 in comparison with OM-1M2 for R and B channels.
For channels R, B in the boost mode, the G9M2 also shows higher DR values compared to the G9. The difference is more than one stop at the base ISO and more than 0.5 EV in the ISO 100 - 3200 range, Plot #9. The difference in DR for the R and G channels between the G9M2 and OM-1M2 is generally not as significant as between the G9, but still for individual ISOs (250, 400) it reaches a value of 1 EV, Plot #10. In case of OM-1M2 the discussed above the “gain-tweaked” improvement in read-noise performance is quite pronounced in ISO range 100 -1000 and 12800-16000.
CONCLUSIONS
Therefore, it can be concluded that, overall, the dynamic range of the G9M2 is significantly higher than that of the G9 and OM-1 over a wide range of ISO values. However, the G9M2 DR enhancement takes place not in all regimes. It is switched off at low shutter speeds (if exposure time is less or equal to one second at ISO100-400) and in case of the burst SH20 mode.
The noise characteristics of the high-gain G2 circuit in G9M2 allow base ISO DR boosting up to 15 stops which probably can be controlled by the firmware.
ON MEASUREMENTS
This section is for those who is interested in details of iWE DR measurements.
The iWE software (v.4.6-01) was used for the measurements, which is available at link:
(file iWE_46-01.zip). At this link you can also download a detailed user manual in English (“Manual_EN_v4X.pdf”), where in section 2.8 ("Data" group) you can get acquainted with how the dynamic range is measured (p.58 -61). You can also watch a short video (file “Data_and_DynamicRange.mkv”) at the link mentioned above.
The image processing in iWE is based on the normalized input data converted to the floating-point format. The normalized data are also used for the statistical calculations. The essence of the normalization is that the maximum value of the pixel response is restricted by a value of 8192 independently on ADC bit-depth. The normalization value 8192 is chosen due to Windows GDI+ 48-bit bitmap format which supports the maximum value 2^13=8192. To make the normalization the ADC data with subtracted black level are multiplied by a normalization coefficient, which is defined by division of 8192 by the maximum ADC value minus a black level value (the division is done in the floating-point format, so no loose in accuracy during the processing).
The response value 8192 not always can be achieved due to non-linearity of the pixel photodiode response at high light intensity. This can happen if one deals, for example, with the extended ISO. For the last reason, in cases the linearity range is unknown, the linearity test must be done. Despite Panasonic does report the linearity range equal to the 16-bit ADC in the EXIF data, so the normalized maximum value 8192 must be achieved, we have performed the linearity test, as shown in Plot#11.
Plot#11. Linearity test for G9M2 at ISO100.
For the linearity test we made white-frame shots. The homogeneously illuminated white paper was shot at the base ISO100 for different shutter speed with increasing the exposure time. We have insured that the normalized output (mean value over 260000 pixels) is linear versus the exposure time and the linearity remains up to the maximum normalized value 8192.
White frames were also used for the gain and photon noise measurements. In this case the shutter speed and lens F-number are fixed, and the response is measured versus the ISO settings, Plot #3.
The Dynamic range is defined as the base two logarithm of the ratio of maximum linear response to the read noise defined as the standard deviation of a pixel values in black-frame data.
This work focuses on the dynamic range of four cameras: G9M2, G9, OM-1M2 and OM-1.
It should be noted that these results were made possible thanks to the tremendous work done by JRSFORUMS, which provided data for the G9M2 not only in the form of black-frames, but also white-frame data for measurements of linearity response on the basic ISO, gain and photon noise for various ISOs. Black-frame data for measuring the dynamic range of the OM-1M2 was kindly provided by LOKATZ. I express my deep gratitude to these people.
DISCUSSION
Let's start by showing the dynamic range measurements of the G9M2 in two modes: i) DR boost mode; ii) SH20 (non-boost) mode, Plot #1.
Plot #1. G9M2 dynamic range in boosted and SH20 (non-boosted) mode.
As one can see, in the boost mode at low ISO values, the DR is more than 2 EV higher than in SH20 mode for all three RGB channels. The boost effect only slightly pronounced above ISO 3200, but small boosting advantage of about 0.3EV remains even at high ISO values up to ISO 25600.
Plot #2 shows the measurements for the green channel of the G9M2 compared to the G9.
Plot #2. DR data for G9M2 in comparison with G9. Both are with ES at 1/32000 s.
As you can see, the dynamic range of the G9M2 at the base ISO 100 reaches a value of 14 EV. This value was tested for three different G9M2 cameras that were available to JRSFORUMS, under different conditions of obtaining black-frames. The reason for this check was that for the black-frames provided earlier by Christian Comes, the measured DR for the green channel is over 15 EV. Perhaps the reason for the differences in the maximum DR values at the base ISO between the current and earlier measurements is due to the changes made by Panasonic in the updated firmware version of the G9M2 camera.
The GM2 value of 14 EV is one stop higher than the value of 13 EV measured for the G9 at the base ISO 200. The dynamic range of the G9M2 exceeds that of the G9, and as you can see from the Plot #2, this excess is at least 0.5EV throughout the all-ISO settings.
It should be noted that the ISO 100-160 range for the G9 is declared by Panasonic as extended, although our measurements show that the analog gain at these ISOs is decreased compared to the gain at the base ISO 200. For example, the analogue G9 gain at ISO 100 is half that of ISO 200. It is also important that the maximum response values at ISO 100 reach the maximum ADC values minus the baseline offset. The only feature of ISO 100, in comparison with ISO 200, is a certain nonlinearity of the response with increasing exposure, which becomes noticeable as the response increases and approaches the maximum possible value. Thus, ISO 100 in G9 can be regarded as the true ISO if the nonlinearity of the response is neglected.
For the G9 at ISO 16000, 20000 and 25600, the measurements showed no change in the analog gain compared to its values at ISO 8000, 10000, 12800, respectively. In Exif, these ISO values (16000 – 25000) are given a double digital gain. Since digital gain always results in a proportional increase in read noise, this fact was taken into account in the data shown in #2 for the G9.
Depending on the ISO, the data for the G9 show the "oscillating" behavior of the measured DR. Since the measurements showed a linear increase in gain with an increase in ISO, this behavior can be explained by the peculiarities of the electronic circuit of the analog amplifier with programmable gains. At certain intermediate gains between the “principal” 200, 400, 800, 1600, 3200, the electric circuit has improved noise characteristics. We will see a similar, "oscillating” behavior below for the OM-1M2 camera as well.
For the G9M2, when switching from ISO 3200 to ISO 4000 (Plot #2), one can see that there is no change in DR. Many sources claim that since ISO3200, Panasonic has applied noise reduction. Our “white-frames” measurements shown in Plot #3 do not support this view. Indeed, at the transition from ISO-3200 to ISO 4000, the read noise does not change, despite the fact that the gain increases in proportion to the growth of the ISO value, and this can be interpreted as some (0.3 EV) noise reduction. However, if we look at the dependence of the total noise (read plus photon noise), where the main contribution is related to the photon noise, we can see that when moving from ISO 3200 to ISO 4000, the noise increases from a value of 117.5 to a value of 148.5. That is, the noise increases by 148.5/117.5 = 1.264 times. At the same time, from data #2, the gain increases from 31.5 to 40.0, that is, 40/31.5 = 1.27 times, which is in very good agreement with the increase in photon noise. Thus, there is no noise reduction for ISO over 3200, and the observed feature should be attributed either to the lower read noise which is due to changing characteristics of the electronic circuit, as in the case of G9, or the features of dynamic range boosting. Similar behavior, when the circuit read-noise performance depends on the gain, can be seen in the example of the OM-1M2 in the ISO ranges 200-800, 800-1000, 10000-12800, see Plot #4 and #5.
Plot #3. Dependence of the ISO100-normalized gain (black line plus symbols), total (red scatters) and read (green scatters) noise vs ISO settings at fixed shutter speed and lens F-number.
Plot #4. DR data for G9M2 in comparison with OM-1M2.
From Plot #4, you can see that the G9M2 has a significant advantage over the OM-1M2 in the ISO range of 100 to 800, which is about one stop (EV). However, for ISO above 12800, there is only very slight difference of about 0.15 EV between the G9M2 and the OM-1, which is comparable with the measurement error estimated as 0.07EV.
It is worth to mention that the OM-1M2 DR is found to be different for the case of electronic and mechanical shutter, Plot #4. Also, it was found that in case of the mechanical shutter the DR for OM-1M2 is very similar to DR for OM-1, as one can see in Plot #5. By the way, it is easy to see from the presented data that at ISO switching in the intervals of 250 - 320, 800-1000, 12800-16000 there are pronounced jumps (growth) of DR, which could be interpreted as the noise reduction by analogy with G9M2 in the range of 3200-4000. However, of course, such an interpretation would be flawed, as shown by the increase in photon noise in case of G9M2.
The high dynamic range mode implemented in the G9M2 does not exist in all regimes, but depends on the low shutter speed and ISO, as shown in Plot #6.
As you can see, the G9M2's boost Dynamic Range mode is implemented throughout the ISO range only for shutter speeds exceeding 1/15 sec. If the ISO does not exceed 500, the corresponding shutter speed is greater than 1 reverse second for the boost DR mode.
Plot #5. DR OM-1M2 vs OM-1 (green channel; mechanical shutter).
Plot #6. Transition (red arrows) to the non-boost DR mode at low shutter speed for G9M2.
Plot #7. The G9M2 boosted DR mode at shutter speed 1/32000s vs the non-boosted modes at an SS of 2 sec and 1/32000 s for the SH20 burst mode.
The boost DR mode is off in SH20 regime, as can be seen in Plot #7. In the low ISO (100 -500) region, the difference in dynamic range between the Enhanced DR mode and the SH20 is greater than 2 EV. In the SH20 mode, the use of two different analog amplification circuits is well pronounced. It can be seen that when switching ISO from 500 to 640, there is a sharp increase in DR by 1 EV, which is explained by the use of the high-gain (G2) electrical circuit with the improved noise performance of the analog amplifier. The low-gain circuit (G1) used in the ISO range 100 -500 allows for true ISO 100, but it is characterized worse read-noise performance compared to the G2 circuit. At low shutter speed (2 s) the non-boost DR performance is different from that in SH20 mode, which again can be associated with some differences in electric circuit, which can be related to different pixel response time necessary in the both cases.
According to SH20-data in Plot #7 the DR at ISO 3200 is just equal to the boost-mode DR. From Plot #6 one can also see that at ISO3200 the DR-jump at 1/15 s is actually absent withing the measurements error. Thus, the noise performance of the G2 circuit at ISO 3200 plays a special role in the DR boosting. Indeed, the data for input-referred noise in Plot #8 clearly demonstrate that the ISO3200 is a “sweet” point at which the input-referred noise is minimal and equal to 0.121. With such a low noise the base-ISO DR in theory can be boosted up to 16 stops by digital dividing the G2 ADC data by a gain of 32. In practice with 12-bit ADC, if the G1 ADC data occupy the highest binary digits of the 16-bit register we have only 4 lowest bits free to work with for the boosting, so the G2-data can be easily divided only by a gain of 16 or lower, so the boosted base ISO DR will be lower than 16 stops.
Plot 8. Input-referred noise measured for the non-boost SH20 mode.
Plot #9. DR for G9M2 in comparison with G9 for R and B channels.
Plot #10. DR for G9M2 in comparison with OM-1M2 for R and B channels.
For channels R, B in the boost mode, the G9M2 also shows higher DR values compared to the G9. The difference is more than one stop at the base ISO and more than 0.5 EV in the ISO 100 - 3200 range, Plot #9. The difference in DR for the R and G channels between the G9M2 and OM-1M2 is generally not as significant as between the G9, but still for individual ISOs (250, 400) it reaches a value of 1 EV, Plot #10. In case of OM-1M2 the discussed above the “gain-tweaked” improvement in read-noise performance is quite pronounced in ISO range 100 -1000 and 12800-16000.
CONCLUSIONS
Therefore, it can be concluded that, overall, the dynamic range of the G9M2 is significantly higher than that of the G9 and OM-1 over a wide range of ISO values. However, the G9M2 DR enhancement takes place not in all regimes. It is switched off at low shutter speeds (if exposure time is less or equal to one second at ISO100-400) and in case of the burst SH20 mode.
The noise characteristics of the high-gain G2 circuit in G9M2 allow base ISO DR boosting up to 15 stops which probably can be controlled by the firmware.
ON MEASUREMENTS
This section is for those who is interested in details of iWE DR measurements.
The iWE software (v.4.6-01) was used for the measurements, which is available at link:
(file iWE_46-01.zip). At this link you can also download a detailed user manual in English (“Manual_EN_v4X.pdf”), where in section 2.8 ("Data" group) you can get acquainted with how the dynamic range is measured (p.58 -61). You can also watch a short video (file “Data_and_DynamicRange.mkv”) at the link mentioned above.
The image processing in iWE is based on the normalized input data converted to the floating-point format. The normalized data are also used for the statistical calculations. The essence of the normalization is that the maximum value of the pixel response is restricted by a value of 8192 independently on ADC bit-depth. The normalization value 8192 is chosen due to Windows GDI+ 48-bit bitmap format which supports the maximum value 2^13=8192. To make the normalization the ADC data with subtracted black level are multiplied by a normalization coefficient, which is defined by division of 8192 by the maximum ADC value minus a black level value (the division is done in the floating-point format, so no loose in accuracy during the processing).
The response value 8192 not always can be achieved due to non-linearity of the pixel photodiode response at high light intensity. This can happen if one deals, for example, with the extended ISO. For the last reason, in cases the linearity range is unknown, the linearity test must be done. Despite Panasonic does report the linearity range equal to the 16-bit ADC in the EXIF data, so the normalized maximum value 8192 must be achieved, we have performed the linearity test, as shown in Plot#11.
Plot#11. Linearity test for G9M2 at ISO100.
For the linearity test we made white-frame shots. The homogeneously illuminated white paper was shot at the base ISO100 for different shutter speed with increasing the exposure time. We have insured that the normalized output (mean value over 260000 pixels) is linear versus the exposure time and the linearity remains up to the maximum normalized value 8192.
White frames were also used for the gain and photon noise measurements. In this case the shutter speed and lens F-number are fixed, and the response is measured versus the ISO settings, Plot #3.
The Dynamic range is defined as the base two logarithm of the ratio of maximum linear response to the read noise defined as the standard deviation of a pixel values in black-frame data.
