duydaniel said:
Hi guys,
Lens resolution is something very often discussed, especially on DXO site.
However, I feel that it is fictional and here is why:
Look at p&s cameras, they often come with low quality zoom lens (compare to DSLR ones)
yet they offer a very high resolution out put images. If you zoom in those images, you can
definitely see details and they are not bad.
If you get the ratio of megapixel from a p&s and the area of light projected on the sensor by the lens and multiply that ratio multiple times until you meet the size of your DSLR sensor, I am sure you can get a much higher lens resolution from your DSLR (not sure if this make sense) than one from DXO.
So if that is the case, then the DXO resolution measurement is a total BS
What do you think?
The problem is that DXO is not actually measuring just lens resolution. DXO's tests are "camera system resolution" tests. The final output resoution of any camera is a
convolution. Lenses don't simply project...they also modify. Sensors don't simply record...they too also modify. The image projected by a lens is then further "convolved" by the sensor, producing a result that has a lower resolution than either the lens (as depicted by its mathematically generated MTF) or the sensor (as determined by it's raw spatial resolution...pixel pitch for raw luminance-only "detail" resolution.)
An example would be to take a *perfect* f/4 lens, one that is diffraction limited and not aberration limited, and thus capable of resolving 173 lp/mm (line pairs per millimeter), and let's say a high resolution P&S camera also capable of 173 lp/mm spatial resolution (which, assuming ideal circumstances and say a foveon-style sensor with 2.8 micron pixels, is plausible). The final output resolution of an image produced by this camera setup will be lower than 173 lp/mm. Because of the fact that the real-world image is convolved into a digital image, we cannot actually achieve that maximum theoretical resolution.
Total "system resolution" is closely approximated by computing the root mean square (RMS) of each component in the system. Technically speaking, each and every specific piece, such as individual lens elements, air pockets, filters in the lens and in the stack above the sensor, as well as the sensor itself, should be taken into account. We can reasonably approximate total system resolution by just taking the RMS of each major component...the lens and the camera. In this case, our "blur circle" size is the same for the lens and the sensor...2.8 microns. The RMS would be:
Code:
tsr = sqrt(2.8^2um + 2.8^2um) = sqrt(7.84um + 7.84um) = sqrt(15.68um) = 3.96um
The pixel pitch and diffraction limited blur of the lens and sensor is 2.8 microns, however the final system blur is 3.95 microns. A larger blur leads to lower resolution. In terms of spatial resolution (lp/mm):
Code:
tsrSpat = 1l / (3.96um / 1000um/mm) / 2l/lp = 126lp/mm
The actual output of a camera setup that uses a PERFECT f/4 lens (something highly unlikely in a P&S, but for demonstration purposes) and a high resolution sensor that uses 2.8 micron pixels, is 126 lp/mm. That is about 73% the maximum potential resolution either the lens or the sensor is capable of (again, assuming ideal circumstances). The ACTUAL resolution of such a setup with real-world parts would still be lower. One can assume that the lens is not perfect...let's say 10% less than perfect. The bayer nature of most CMOS image sensors will also impact results, possibly shaving off another 25%. That would give us a lens capable of 155lp/mm, and a sensor capable of 130lp/mm. That results in a total system resolution of 100lp/mm.
Now, keep in mind...this is for a camera with a very good lens, and a sensor with much smaller pixels than the average DSLR. If we run the numbers for, say, the 5D III and a high end L-series lens at f/4, we get even lower results. Assuming we get the same 155lp/mm out of the lens, the sensor now has 6.25 micron pixels. Assuming the same 25% loss due to the bayer array on the sensor, and the final system resolution comes out to 57lp/mm.
Before DXO moved to their new way of presenting lens tests, most of the higher end L-series lenses were getting 40-55lp/mm resolution results with cameras like the 5D II, 5D III, and 1D IV. That is right in line with the math I've demonstrated above. My 10% and 25% weights for lens and sensor are a bit optimistic and rather simplistic, one will usually experience greater losses, further diminishing resolution.
Additionally, one has to understand DXO's test setup. They test lenses in extremely low light, much lower than the kind of light people will photograph in. This has the effect of severely hurting any lenses that are not ultra fast (i.e. f/2 or faster) because the "transmission" factor takes a serious hit. The only logical conclusion I can come to is that to maintain a "consistent" test setup (consistent by DXO standards anyway), a lower light level had to be used in order to test ultra fast lenses (i.e. f/1.2 or faster) without over-exposing. Again, this is similar to DXO's other tests, and one particular factor has an overbearing impact on the total "score" of the lens in the end...and undue impact, and one that does not fairly or appropriately demonstrate the true quality of many lenses. (An example of that would be Canon's new 200-400 f/4 1.4x TC lens...which scores a mere 24 on the DXO scale, when it is a superb lens with excellent resolution and excellent characteristics in general. The SOLE detractor? TRANSMISSION!!!)
(NOTE: I would actually offer that DXO's lens tests are fundamentally flawed when it comes to their transmission factor. I do not believe they are actually testing "light transmission"...if they were, then it shouldn't matter what the aperture is...all that would really matter is how much of a known quantity of a small column of light transmitted through the center of the lens actually reached the back of the lens. Technically speaking, such a test should work even at the minimum aperture, as aperture should not play a role in testing the light transmission of a lens at all...however in the case of DXO's tests, it clearly does.)