A Brief M4/3 Statistical Sensor Classification Between Panasonic and Olympus in Combination with Cropped F. and Full F. Has M4/3 Reached Its Limits?

 Attention to MF3 sensor has been picking up since it’s introduction to the market, especially in the recent years with it’s improvements on it’s dynamic range to make its gap tighter with higher quality sensors such as cropped sensor. Therefore it is logical to compare the most recent MF3 sensors with the cropped sensors.  I have decided to do PCA analysis with the inclusion of MF3 sensor dataset to the previous Canon, Nikon and Sony sensor dataset to get a better overview of the position of MF3 sensors in a bigger picture and where they are standing with regards to cropped and full frame. The RAW sensor data used in this article is extracted  from photonstophotos.net as mentioned in the previous article. I have received multiple emails about the readability issues on the a post and I hope the current figures are improved enough for those audience.

The following statistical analysis test and interpretation on two major Micro Four Third (MF3) rival brands is based on the previous post I published a while ago that I encourage the audience to read it,  therefore I will not reiterate the methodology here and I only focus mostly on explaining the results of this test.  For this test, nine sensor variables’ data from selected cameras were used as previously listed . The selected MF3 camera sensors are shown in Table – 2. Olympus E3 sensor parameters’ data is also added in order to present give an extended temporal span to the evolution of MF3 formfactor throughout the years.

Table-1

Camera list.jpg

Table -2  MF3 Sensor Cameras included in to do a new PCA to the available list of Canon, Nikon and Sony

Using the multivariate test of principal component analysis (PCA) with the available dataset from 9 sensor variables RAW values (Table -1 ) that were acquired at their base ISO from Photonstophotos.net  , presented quite a distinct separation between MF3 vs Cropped F. and Full F. sensors (Figure -1) . The PCA on the sensor data presented in Figure 1 shows that the two PCs  (PC1 = x and PC2 = y axis) explained a cumulative of 69% of variability in the sensor data from the listed camera models. That is a large portion of the information in the data; therefore the sensor relationships can be interpreted with a high degree of certainty.

MF3_FF_CF_Updated.jpg

Figure – 1  PCA showing clustering of three sensors form factors from five camera brands of Canon, Nikon, Olympus, Panasonic and Sony. The distance of QE, LLEV, PDR, Pitch, Read Noise FWC, Sensor Size ,DSNU and PRNU from the plot origin (centre) also shows the amount of leverage these variables have on the distribution of the data points on PC1 and PC2 axes.  Download the high res figure from here

The correlation and significance of the mentioned nine variables and their influence on the PC data structure (Figure 1) were examined with a variable loadings plot and are presented in Figure 2 which all loading vectors run between -1 and +1.  The correlation loading plot represents the leverage and the influence of each sensor parameter on clustering of sensor data. Overlaying this plot that is schematically shown in Figure 1 will explain by which parameters or features each camera or a group of cameras are best explained with or why they are grouped or clustered together.

Correlation loadings.png

Figure 2 – Correlation loadings  (correlation coefficient) of sensor parameters.  Diagonally located variables are anti-correlated (inverse correlation). Closely located variables are highly correlated (they co-vary). Inner and outer ellipses explain 50% and 100% of the variance (r2 =0.5 and 1).

The correlation loading plot represents the leverage and the influence of each sensor parameter on clustering of sensor data. A schematic overlay of the loadings on the data is shown in Figure 1 which helps to understand by which sensor features each camera or a cluster of cameras are best explained with or why they are clustered together.

As an example if we focus on MF3 cluster (purple oval), it can clearly be observed that the Olympus has made a distinctive departure on its EM-1 Mark II sensor characteristics mostly on its QE comparing to its own and the rest of Panasonic MF3 sensors. The QE is noticeably improved on this camera which can indicate as an attempt to reduce read noise and compensate for the limiting MF3 sensor size factor. This can also be seen on figure 2 where Read Noise generally runs countercurrent with QE. To prove this, the distance between QE and Read Noise loadings is larger than other MF3 sensors for EM-1 Mark II. This indicates that the QE leverage is more pronounced on Olympus EM-1 Mark II and therefore demonstrates less noise on it simage than Panasonic Lumix DC-GH5 according to the PCA results.

YearsOverlapped

Figure 3 –  Temporal trend showing how the changes in sensor attributes developed throughout years of sensor manufacture (see also Figure – 1 and Figure 3a for actual camera names for each data point).  Download the high res figure from here

ZoomedYears.jpg

Figure 3a – Blow up of Figure 3 on MF3 camera clusters (lighter oval). The suggested MF3 and Cropped F. areas are proximate. Download the high res figure from here

CMOSCCD_etcUpdated.jpg

Figure 3b – PCA score plot on sensor type comparison. Download the high res figure from here

The previous post’s PCA coordinate system holds also for the current PCA in figure 3a in this article; that means the more a camera score shifts from PC1 (x axis) zero value towards negative values (left side of the plot), the less the quality of the sensor image will be.

As can be seen in Figures 3a and 3b there are some apparent overlaps from Cropped Frame cameras on MF3 camera groups’ region such as Canon EOS 7D Mark II and Canon EOS 80D which clearly shows the similarity of their sensor characteristics to that of  certain MF3 form factor such as Panasonic Lumix DMC-GH4 and Lumix DMC-GX8 respectively. A close sensor similarity can also be seen between Canon M3 (as a cropped sensor) with Olympus OM-D EM-10 with almost similar noise attributes according to the PCA scores.  Despite earlier production year of Canon M3 (2015), the M3 produces less noise comparing to its newer sisters; Canon M5 (2016) and Canon M6 (2017). Needless to say that  by looking at the score plots in figure 3a, all three Olympus EM-1 Mark II, Panasonic Lumix DC-GH5 and Olympus Pen-F image characteristics are superior to the above mentioned Cropped Frame Canon sensors.

Dendogram.jpg

Figure 4 –  A dissect from the entire correlation tree of camera sensors where MF3 cameras are located. Cameras are similar by their relative distance; the less distant their branches are to each other the more similar are their sensor profile. Note the different classification of EM-1 Mark II with regards to the rest of MF3 group, it has almost unique sensor attributes comparing to other groups of nearby sensors. Download the high res figure from here

Best MF3 camera sensors

Certainly the sensore scores from this statistical analysis are sitting quite close together (indeed a close competition)  however  the outcome of the sensor score plot (Figure 3a) implies that the Olympus EM-1 Mark II appears to have the most improved MF3 sensor which has slightly a better characteristic over Panasonic Lumix DC-GH5 and Olympus Pen-F  as the second and third place respectively EM-1 Mark II  also shows less noise due to its improved QE.

Conclusion

Clearly MF3 manufacturers, as well as Cropped F. and Full F.  are pushing towards achieving lower noise and improved dynamic range as shown in my previous post. The recent years sensor scores’ trend are all pushing toward the right side of the PCA plot (see Figure 3). Since MF3 has its specific sensor size it would appear that MF3 is reaching to some technical limits and therefore it would be interesting to see how this problem would be worked around. OLympus has pushed far on QE to maintain superiority; but what is the next step? Is the market going to see back illuminated MF3 sensors?

Acknowledgment

I would like to thank Mr. William J. Claff for his permission on using his Heatmaps and other data for this work from Photonstophotos.net . Once again the outcome of this statistical analysis test with the inclusion of additional sensor data demonstrates that the measured Heatmap data derived from  Photonstophotos.net are reliable.

Emmett E. Rad  / February 2018

Please support this website if you find my articles informative
Donate with PayPal

Advertisements


Categories: Uncategorized

Tags: , , , , , , , , , ,

%d bloggers like this: