GH1 vs GH4 - generations

well, having compared my GH1 some time back to an OMD E-M5 (here) and compared Out Of Camera JPEG images and then some post processed RAW images from both I came to the conclusion that the cameras were about a par on image quality and that there was little or no compelling reason for a stills photographer to upgrade their GH1 in that direction on the basis of image quality if they are inclined to shoot raw.

Since we just got a new GH4 in our office in the Video Production unit I thought I'd have a look at the two cameras side by side to see what has really changed. Again  this perspective is from that of stills (which such cameras are validly used for).

premise of comparison

I am fundamentally a stills photographer. Further, I know my craft and I don't feel any need for (nor do I actually like working with) bucket loads of automation. I find it distracting and takes my thoughts away from the actual job of making an image. I hate fiddling with multiple settings and normally work in Aperture priority or Program Modes. I shoot in RAW so that I can process afterwards because I've never been fully satisfied with OOC JPEG for everything.

So if that does not sound like you and you are into video then its simple - the feature sets of the GH4 make it streets ahead. Just get it.

So for me I had some questions to answer, and some were already clear in my mind. What was clear?

Specs: I knew that the image sizes of the cameras were 12Megapixel vs 16Megapixel ... that would be clear. All cameras that do in camera image correction (or later in software such as LightRoom if you wish it) absorb a number of the pixels in that correction. So while the uncorrected width of the GH1 is 4016 pixels (not 4000) the the same is true for GH4 which goes from an uncorrected RAW pixels of 4624 down to an OOC JPG of 4608.

However its also mysterious to people how to interpret pixels, they fail to grasp that its actually a square ratio and that to understand the comparison we should take the square root.
For the GH1 the square root of 12Mp is 3.46 and for the GH4 at 16Mp that's 4. Suddenly the numbers don't look so importantly different.

Lets put that number set into a visual. So if you print to the maximum DPI of the cameras then this is the difference in print size you'll gain...

Didn't blow my sox off either ...

But I also knew that the GH1 had something that the GH4 had foregone, and that is the multi format sensor (discussed here).  Meaning that for anyone who shoots stills in 16:9 that the GH1 will have a bit more up its RAW sleeve with 4368 pixels wide. This reduces the gap between the two sensors in some situations. You can see clearly that the 4:3 records a greater height than anything and the 16:9 records a greater width. Looking in particular at the 16:9 compared to the 4:3 we see this:
So the 16:9 is recording 352 more pixels (and importantly) by actually capturing a wider area of the image circle cast by the lens.

Lastly I do not believe there has been any real (meaning x2 or better) improvement in sensors since 2010. Sure, pixel pitch has come down, and sure read times are faster ... but to a stills photographer its really only

• the actual shutter you get 
• for the actual aperture you've set
• at the actual ISO you've set, 
• and the signal to noise

that's the big ticket items to me.

The Nitty Gritty

All shots were taken with the same lens (Panasonic 25mm f1.7) at f4 to minimise any focus issues in the area. I then took shots with the camera on Av and just bumped up the ISO from base to max in 1 EV steps.

Ok, so speaking of ISO, lets look at the ISO differences, the GH1 was "under rated" at birth to have a ISO100 that was effectively measured at 137 anyway, giving you an advantage which continues to be almost one stop throughout the range. ... This is made clear in the DxO measurement chart (see here)(and my testing showed shutter speeds which were consistent with this point)

so you can see that when ISO 400 is selected  on both cameras the GH1 gives 591 while the GH4 gives shutter speeds that would match ISO 227; which is about a full stop difference (a little more actually) in favor of the GH1.

The result of this (should it need to be made more obvious) is that when thinking of image noise you need to compare the GH1 to the setting above it on the GH4 to be equal.

Yes, that means at 1600 ISO the GH1 is essentially giving you sensitivity (and thus shutter speeds for a given aperture) of the GH4 set to 3200. Meaning that the GH4 matches the normal King Wang user who looks only at numbers and not at results.

Bit Range

Recently I looked at the Sony A7 using a RAW file analysis method (see here) that allowed me to see exactly where the data started recording (on which digital quanta {see quantisation if confused}) and where it ended. From this I can see clearly which camera has what range of data.

NOTE: it has been observed by a reader that I may be mistaken in this approach and that the true issue is the voltage output range not the quantised data range as a reflection of white and black points. That may be true (or may not be if Panasonic did not specifically recalibrate their sensors to their ADC), but not necessarily. If you are seeing more in the dark than me, or more in the highlights than me then great ... but to me I see about the same. I'd need to use a Stouffer Step Wedge to be sure of the exact ranges (which I didn't think of). 

So lets have a look at that. First here is my test shot setup.

I have (in the red circles) everything from clipping to dark and murky  ... specifically to test the range of each. The red square is the point of focus.

Shots taken on a tripod and processed variously with my tool of choice DCRAW ... why dcraw? Well for a start its an equaliser, it shows things without "fix ups" in software (which can be done to both if you like) to allow me to see the real RAW data not the make up department and touch up department: not soRAW.

So the GH1 consistently started recording pixels at quanta 16, while the GH4 started recording at quanta 142 ... both started recording at the same time in all channels (which is nicer if you ask me and nicer than the A7). Clearly the GH4 is ignoring some 'low level' quanta which would probably be to avoid "floor noise".

Then where does it end? Well the GH1 decided to put the last of the clipping in at quanta 3990 while the GH4 ran out to 4097. Which is interesting as this gives the GH1 an actual recorded data range of 3974 and the GH4 slightly less at 3953 levels.

So:

• Start = 16 vs 142 (earlier in the range than GH1)
• End = 3990 vs 4097 (the GH4 records longer into the range ... but)
• range = 3974 vs 3953 (actually less!)

So it is interesting that the DxO ratings suggest that the GH4 has a greater bit depth (23.2 bits vs 21.6 bits). I don't see how this works as bit depth is the range.

Which isn't what I find ... So to me the GH1 has actually a slight lead on the amount of bit depth available over that of the GH4. This suggests to me that DxO is only looking at the data end points or that they think there is more noise in the GH1 low end. I'm not sure that this the case here, and I don't know how they measure it (but I do know what they say is how they calculate it). This matches the findings that I had with the A7 in my earlier post (or they are doing something more cunning).

So what does the Noise look like

Well they were all good to the 1600 setting on the GH1. So this is ISO 1600 on the GH1 and 3200 on the GH4 , given that these settings mean that the actual ISO for each was 2154 and 1860 respectively. Yes, that's right the GH4 when set to 3200 actually yeilds a wee bit over 1600 but the GH1 yeilds almost 3200. This is reflected in the shutter speeds which were 160 and 200th respectively.

since these are both 100% pixel views you will see that the above chart showing relative print sizes is about bang on.

So, Noise looks startingly similar (as one would expect if there has indeed been no progress in sensors as I expected).

Next lets look at the maximum ISO of the GH1 which is 3200 ... according to DxO that works out to be a measured ISO of 4176 for the GH1 and with the GH4 set to 6400 gives a measured ISO of 3835, again lower despite being higher.

This time we see that the GH1 is falling apart more than the GH4 is in two ways:

• a colour shift is occurring (I did dcraw with -w to use the cameras white balance settings on all images)
• and noise is starting to become striated and follow patterns, while the GH4 is still looking "organic" in its noise.

So myself I'm going to say that I'd be happy with the GH4 at that higher ISO much more so than the GH1 ... which fits what I've always found that one ISO stop below MAX ISO is the limit effective limit.

So, lets look at what the GH4 gave up beyond the GH1 limit of 3200

well its dirtier that's for sure ... is it usable to you? Sure ain't what I'd want to use.

Lastly lets step outside of the box and compare OOC JPEG from the GH4 @ ISO 6400 with the GH1 @ ISO 3200 made with RawTherapee (a modern but free editor)

This is a central portion of the image at 100%

So GH4 ...

which looks clean and sweet ... but shows (to me) strong smudging which is the hallmark of more aggressive noise reduction.

GH1...

with only primitive noise reduction.

Less clean for sure, but you know, I can read those book ends better (meaning its sharper)

afterthought

after writing this I came across the NIK suite for free (thanks Google) and thought I'd try applying some noise reduction to the TIFF from the GH4 (and nothing to the GH1) file. So this is the result of that

To me the GH1 is a usable result while the GH4 is still rough. Sure you gain a shutter speed increase from to 320th up to 1600th of a sec (two stops as lens was f4 in both) but I would wonder (now that I think about it) would the GH1 have been more like that if I'd just pushed it 2 stops by increasing the shutter speed and compensating in post?

dunno ... something to try next...

conclusion

So there you have it, some interesting questions raised (well interesting to me) about how DxO measurements should be understood (clearly its not that simple) and some results that support that for a stills photographer that the GH1 is still as good a camera as the day it was released, maybe even better than almost anything around today (as mentioned it was hardly less than the E-M5 I pitted it against). Considering you can get the GH1 for under $200 its got to be a consideration.

The GH1 also brings some interesting features to the table for those who may wish to shoot other than 4/3 ratio and still makes great images. That your wide lens becomes a little wider in 16:9 than any other camera has to count for something (especially given that it comes with no penalty).

It supports my views that the major benefits of the current cameras are:

• that you can buy them new (needed for some businesses)
• have better in camera features

If you're a budget conscious shooter and your focus is on the produced images then a used GH1 is a great buy, you get:

• flip screen (which people now are discovering how handy it is)
• great EVF, among the biggest even today
• great image quality, equal to almost anything in m43 today
• lighter and smaller
• great battery life
• access to cheap batteries

probably its the "sleeper cell" killer camera of the m43 world ... still ...

Testing LS-4000 with Stouffer Stepwedge

Now that I have a LS-4000 in my home I thought I would apply some of the same testing that I have previously used on my Epson flatbed scanners. I used the FH-3 holder to hold a Stouffer Stepwedge and scanned it in sections to see how the scanner responded to levels of denisty.

I was stunned how it ran out of grunt in ability to penetrate the dark areas. I noticed how sensitive to the 'clear' area of the wedge it was so I popped a little bit of analog gain into the scan (0.8 actually). I did this to hopefully extend the dynamic range into the dark areas and bring the white areas up to begin clipping at "film base".

Cutting straight to the results it is not only not good, its actually surprisingly not good.

This is the result of a scan with 0.8 of analog gain applied to the master gain.

I have plotted a Log₁₀ graph of the results in red (as this is how scanner responce needs to be measured for photographic purposes).

It seems to have made a strange change to the "linearity" of the scanner with a marked step occurring at about step 13.

Its strange that both parts remain more or less straight with a deviation in the middle.

Further, the 'blooming' of the scanner is apparent in the darker areas (note the small intrusion of one of the "frame" supports in the FH-3, this shows that the depth of black is not fully reached even though the ability to discern graduations is substantially impaired).



out of interest this is the result from my Epson 3200, and no blooming is evident in this.



so ... this makes the result from the Nikon particularly scary. Feel free to compare this to the results from my Epson 4990 scanner here.

Note: this effect of blooming is also apparent with x8 oversampling, so it is not just a digital noise issue.

I also thought it was worth mentioning that not only is there substantial bloom (clearly visible here) but that altering the analog gain disturbed the levels of RGB too. Looking at the RHS of this graph you can see three peaks, this is the places where Red Green and Blue have moved up towards 255 but clearly unevenly.

So, without careful profiling using the analog gain on this scanner to 'punch into dark areas' will result in strange colour shifts. I advise use with caution.

I left this section 'zoomed in' to show just how much blooming results in the transition from dark area to light area, it is significant.

---

Addendum

I was asked some additional quesitons about this so I have repeated the tests and included results for 0 analog gain.

NOTE: rather than wait over the weekend I've disassembled the scanner and sure enough the mirror was rather filthy. So a quick clean with windex and a cotton bud (removing the mirror of course) later and its now performing well


My method was to scan with the scale set as linear from 0 to 255, the scanner software is set to put the data into Bruce RGB, so after scanning this profile was assigned to the data to ensure that the results are correct. As the wedge is so long, I had to scan it in 3 sections. To perform this I needed to

• insert the wedge strip into the FH-3 holder
• disable any preview or focus on insert
• scan each step without prescan or preview

some overlap was possible in scanning (as it is manually fed) so I can confirm that (for instance) the values of step 4 (and the clear part of the step number) were consistent from scan to scan.



The results (data first) include the stand deviation of the values in the selection swatch. Standard error is calculated as the percentage of standard deviation to the median value (not, I use median in this testing not mean).


It seems to indicate that the error is getting quite high as the values go above 13 on the stepwedge. This needs to be taken into account when assessing how effective alteration of the image with curves will effect colour accuracy or noise in shadow details. Now the graphical view




seems that the linearity step at about 13 is still there

Epson 4990 response testing

Some time ago I tested my Epson 3200 scanner for its ability to plumb the depths of a Stouffer wedge. The reason is to understand (not guess at) how much density the scanner can cope with (meaning how much shadow detail can be recovered from slide scanning).

I scanned the wedge as a linear full range positive and then input the median values obtained from each 'patch'. That's the blue line is the scan value (255 - 0) returned, the red line is the Log₁₀ of that.

Looking at the straight numerical levels shows that it really starts to plateau out at about 17 on the step wedge. It seems like a good result, but when compared to the much older 3200 its not that much different.


Since it is the log difference between the values that is important in film scanning I've placed them in the graph. Interestingly the 3200 actually remains linear in the log responce longer than the 4990.

I don't know if this is a good or a bad thing as film already has non-linear responses in the shoulder of the curve anyway.

Well, anyway it seems that the 4990 goes into the shadows a little better than the 3200, but despite how it seems on scanning film its not screamingly better when measured.

Using Epson flatbed scanners for densitometery

In his book "Beyond the Zone system" Phil Davis outlines how to adapt a pentax spot meter to be used as a make shift densitometer (for checking negative density in understanding development and traditional enlarger printing). Well, if that can be done (and calibrated) why not use a scanner with a known (and constant) light soruce which already seems to be calibrated. I this article I explore using my Epson in just that way.

Two of the scanners I use for film scanning are Epson scanners (a 3200 and a 4870). These are not exactly cutting edge material, but then when scanning 4x5 inch sheet film maybe they don't really need to be. In terms of resolution, they do not cut the mustard (so to speak) when scanning 35mm with the intention of printing bigger than 5x7 inch prints, where a Nikon LS-40ED or LS-V ED is better, but its easy to get 1200 dpi out of them and some bureaus only provide 2400dpi for A0 sized prints anyway.

Anyway, back to the Epson. I've been interested to know just how well it really performs, so I bought a Stouffer stepwedge and scanned it to see how the scanner would perform. This is the result of a 'linear' scan with no curves applied.


Stouffer publish the data on this step wedge as:

number of steps	density change	density change in stops	Dmax
21	0.15	1/2	3.05

The Epson scanners have a 'calibration' area at the top of the plastic film holder. This is used internally in the scanner to (I assume) to reference check the light source. I reasoned that if I covered that with something that blocked some of that light then I might be able to get better black out of it.

So, I got a bit of green OHP transparency sheet and used that to cover the scanner in a series of scans. I scanned the step wedge in the following manner:

1. 16bit RGB scan
2. 16bit greyscale
3. 16bit greyscale (one sheet of green OHP transparency)
4. 16bit greyscale (two sheets of green OHP transparency)

My results are interesting.

tablet step	RGB no cover (Green channel)	greyscale plain	greyscale 1 sheet	greyscale 2 cover
1	230	233	254	253
2	192	192	226	237
3	153	155	183	210
4	126	127	150	175
5	101	104	123	142
6	85	85	102	120
7	69	71	83	96
8	56	57	67	79
9	45	47	56	65
10	38	40	46	54
11	32	33	38	45
12	26	28	33	38
13	24	24	28	32
14	19	20	24	28
15	16	17	20	24
16	15	15	18	21
17	13	13	16	19
18	11	12	15	18
19	10	11	14	16
20	10	11	13	15
21	9	10	12	15

Here is the graphic view:

You can see that a log(10) plot (scale on the LHS is more or less linear until about step 17). The straight sample data is plotted on the RHS of the graph.

My interpretation of this is that:

1) the scanner runs out of grunt at about step 16 or 17 and only pulls the last dregs out of shadows. I don't mind this for my negatives (which are seldom so dense), but I think it makes it clear that this is not the beastie for scanning slides (especially slightly dense ones).

2) covering the calibration area yields better white point selection, with the best results (not published here) coming from using a piece of unexposed but developed film as the cover (say a section of film leader, or sacrifice a sheet of film). This has the advantage of maximizing the separation of the data and bringing up the dark end of the scale by a stop or so.

I will be keen to test my 4870 next