I agree with Bob S., you've painted yourself into a corner, Ciprian.
One ballpark indicator of technically acceptable recordings is the
histogram. If it doesn't pile up on the edges, it's basically okay because
there is information in the deepest shadows and/or brightest highlights.
When editing, make sure you don't clip it. Beyond that, play.
As Paul says, your eye is a far better judge of quality than software. And
just like AIs or neural networks, it has to be trained by being presented to
a lot of examples. The analogy probably doesn't hold much further than that,
but you'll do good to indulge yourself. Leisurely.
best,
Jostein
-----Opprinnelig melding-----
Fra: Ciprian Dorin Craciun
Dato: 30. desember 2013 17:13
Til: pdml@pdml.net
Emne: Regarding the assessment of raw image technical qualities (i.e.correct
exposure, focus, sharpness, etc.)
Hello all!
Before describing the problem, please let me stress that I know
very well that the "quality" of a photograph lies more in its
aesthetic properties than in its technical ones. However in this
thread I don't discuss about "photographs", but "images", i.e. the raw
data (pixel values) that are totally decoupled with what they
represent. (Thus I hope I won't start a flame-war regarding
aesthetics vs. technical qualities.)
(If you want to skip the "context" and go directly to the problem,
just go to the section titled `The questions`.)
== The context ==
So the problem, I always capture "images" --- note how I didn't
say "photographs" :) --- in RAW format. And, while working in fully
manual mode, very often I obtain multiple "images" of the same scene,
where the main variable is the exposure time. Also while in low
light, either interior or at night, I strive to use a low ISO like
200-400 thus with an exposure time of about 1/10 of a second, and
although the SR is enabled, I always use continuous drive (in low
mode) to take about three exposures, in the hope that one of them is
sharper.
The problem? During the editing process, after I select which
"scenes" are the ones I like the best, I end up with a lot of "images"
of the same scene (with almost identical composition). And thus my
problem is which of the two or three "images" is the most suitable
one, in terms of technical properties, to be used as input for
processing (thus obtaining in the end my "photograph").
== The questions ==
Thus my questions are the following:
(A) Which are the technical qualities I should look for in the image?
(B) What software tools exist out there that would help in the
assessment of these technical qualities?
== My answers ==
For (A), the technical qualities, I've identified the following:
* Is the subject properly focused? I.e. since I use auto-focus,
sometimes with slim subjects it might happen that the camera focuses
on the background.
* Is the image sharp "enough"? I.e. not blurry due to shake,
miss-focus (back or front), subject movement, etc.
* Is the exposure "correct" / "optimal"? I.e. especially since I
use ETTR (expose-to-the-right):
* did I overexposed something, and if so which channels, which
parts of the picture, where are the "next" shadows relative to the
highlights, etc.;
* do I have underexposed areas, and like above how much, etc.
Did I miss other technical properties?
I've purposely dismissed the following:
* noise -- which by its nature is a property of the sensor and
constant in effect for a given ISO value;
* depth-of-field -- which manifests itself as a lack of sharpness
for the objects outside the DOF; (although it would be nice to be
able to estimate how far a given pixel is outside the DOF;)
* flare -- they are easily discerned in the image;
* bokeh, vignetting, distortion, acutance, etc. -- which are
invariable properties of the lens (for a given aperture and focal
length); (and bokeh is a highly subjective quality;)
* chromatic aberrations -- like above, although would be nice to
highlight them;
Unfortunately for (B), how to asses these qualities, I have a few
theories (see the next section titled `The science-fiction`) and less
ready-to-use solutions...
First the solutions. Currently I'm using Geeqie (an Linux-based
image manager) to find the scenes I like, and for each scene which are
the obvious failed images, i.e. those with clear over- or
under-exposure or blur, and I end up with a few candidates. However
even if I zoom 100% I don't see the raw image, but the embedded JPEG,
which was already processed by the camera, thus I can't correctly
assess the sharpness nor the exact exposure "optimality" (due to white
balance).
Thus I then switch to RawTherapee --- which I've configured so
that by default it doesn't apply any settings at all, just neutral or
disabled values --- to look at the raw channels histogram, the exact
image without added sharpening, and the focus mask. However I can't
use the "blinkers" (i.e. underexposure / overexposure masks) to see
which parts of the image are burned, because they work after white
balance. The focus mask shows me which points are in focus, but give
no quantitative feedback of the sharpness. Moreover switching from
one image to another takes for ever.
I know there are a few applications like Rawshack, which give me a
numerical analysis of the raw data, thus helping me assess the
optimality of the exposure. However there is no visual feedback, and
no solution for the other qualities like focus or sharpness.
Did I miss some other tool?
== The science-fiction ==
However, because by trade I'm a developer, I've started pondering
about a do-it-yourself solution. My theories --- which I've started
to put into practice with a small prototype --- revolve around:
* For assessing exposure it is clear: I take the raw image, split
it into the four channels, one for each of the 2x2 Bayer mask --- thus
without any interpolation, and at half the sensor resolution --- and
then posterize each channel with only 6 (or more) levels: two for over
/ under exposure, and four (or more) levels spanning the rest of the
exposure range, approximately 2 EV apart (they can be seen similar to
the "zone" concept). I then can display these channels individually
(or recombined in a gray image) to see which levels are used (and how
much). I get thus a kind of an "exposure map", just like a heat-map
in scientific applications.
* For assessing both the focus and sharpness: I start from the
same channels, and apply a high-pass filter, which after normalization
I multiply with the original channels (thus the result of the filter
acts like an opacity / alpha channel for the original image). Thus I
obtain an image where the in-focus parts are brighter and out-of-focus
parts are dimmer.
* In all the operations above I apply various iterations of a
dilation filter, which helps to better visualize the areas where the
problem lies. (Although it reduces accuracy it increases visual
feedback.)
I've been playing with a prototype that implements the above
ideas, and I can submit some samples if there are interested people.
(Currently a lot is hard-coded, and builds only on Linux.) :)
Does it sounds crazy? :) Does someone have some ideas I could try
out? Can one, more scientifically-inclined, point me to some papers
or similar research?
Happy new year! (And Sorry for the lengthy email...)
Ciprian.
--
PDML Pentax-Discuss Mail List
PDML@pdml.net
http://pdml.net/mailman/listinfo/pdml_pdml.net
to UNSUBSCRIBE from the PDML, please visit the link directly above and
follow the directions.
--
PDML Pentax-Discuss Mail List
PDML@pdml.net
http://pdml.net/mailman/listinfo/pdml_pdml.net
to UNSUBSCRIBE from the PDML, please visit the link directly above and follow
the directions.
