[FFmpeg-user] Question about "normalize" filter
Michael Koch
astroelectronic at t-online.de
Mon Jan 30 01:22:46 EET 2023
Am 29.01.2023 um 23:36 schrieb Paul B Mahol:
> On 1/29/23, Michael Koch <astroelectronic at t-online.de> wrote:
>> Am 29.01.2023 um 23:07 schrieb Paul B Mahol:
>>> On 1/29/23, Michael Koch <astroelectronic at t-online.de> wrote:
>>>> Am 29.01.2023 um 22:05 schrieb Paul B Mahol:
>>>>> On 1/29/23, Michael Koch <astroelectronic at t-online.de> wrote:
>>>>>> Am 29.01.2023 um 19:32 schrieb Paul B Mahol:
>>>>>>> On 1/29/23, Michael Koch <astroelectronic at t-online.de> wrote:
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> if I understood the documentation correctly, the normalize filter
>>>>>>>> maps
>>>>>>>> the darkest input pixel to blackpt and the brightest input pixel to
>>>>>>>> whitept:
>>>>>>>> darkest pixel --> blackpt
>>>>>>>> brightest pixel --> whitept
>>>>>>>>
>>>>>>>> However I need a slightly different mapping:
>>>>>>>> A black input pixel shall remain black, and the brightest input
>>>>>>>> pixel
>>>>>>>> shall become white.
>>>>>>>> black --> blackpt
>>>>>>>> brightest pixel --> whitept
>>>>>>>>
>>>>>>>> With other words: Just multiply all pixels by a suitable constant.
>>>>>>>> Don't
>>>>>>>> add or subtract anything.
>>>>>>>> Is this possible?
>>>>>>>>
>>>>>>>> Known workaround: Make sure that the input frame contains a black
>>>>>>>> pixel,
>>>>>>>> by inserting one in a corner.
>>>>>>> Try attached patch.
>>>>>> How must I set the options for the desired behaviour?
>>>>> Set first strength to reverse of second strength.
>>>>> So 1.0 and 0.0 or 0.0 and 1.0
>>>> I did try with strength=0:strength2=1 but the output isn't as expected.
>>>>
>>>> I'm using this input image:
>>>> http://www.astro-electronic.de/flat.png
>>>>
>>>> The pixel values are about 171 in the center and 107 in the top right
>>>> corner.
>>>> The center to corner ratio is 171 / 107 = 1.6
>>>>
>>>> In the output image I measure 248 in the center (which is almost as
>>>> expected, probably correct because I'm measuring the average of a 7x7
>>>> neighborhood), but I measure 122 in the top right corner.
>>>> The center to corner ratio is 248 / 122 = 2.03
>>>> The corner is too dark.
>>>>
>>> I checked with oscilloscope filter (s=1:tw=1:t=1:x=0), far left pixels
>>> (as they are darkest) and they are not changing (min values are same
>>> with and without filter run)
>>> With default parameters and just strength(2) set to your values, so
>>> the darkest pixels are left untouched. Did not checked brightest
>>> pixels output, but they should be correct too.
>> But that's not the behaviour I need. All pixels shall be multiplied by
>> the same suitable constant, so that the brightest pixel becomes white.
>>
>> Input center: 171
>> Input corner: 107
>>
>> constant c = 255 / 171 =1.49
>> Output center: 171 * c = 255
>> Output corner: 107 * c = 160
>>
> Normalization does not do that and that functionality does not belong
> to such filter, it stretches ranges of all pixel values so they reach
> maximal possible range.
Can't you just add an option that disables the minimum finding
algorithm, and set the minimum to zero (=black)? That would do the job.
Let me explain why I need this kind of normalization.
Most lenses have vignetting. Especially in astronomical images,
vignetting must be corrected before any other image processing can be
done. For this purpose a flatfield image is taken with the same lens at
the same aperture, but with a uniform white screen in front of the lens.
Normally the flatfield image is exposed at roughly 50% gray level, to
avoid problems with nonlinearity near white level. The linked image is
the flatfield.
Vignetting in an astronomical image is corrected by dividing the image
by the flatfield. That can be done with the blend filter.
If I use the flatfied as-is (with roughly 50% gray level), then the
image is effectively multiplied by a factor 2. That's a problem because
bright pixels might get clipped. To minimize this problem, I want to
normalize the flatfield as close to white level as possible, before
using it.
Michael
More information about the ffmpeg-user
mailing list