Vignetting, discussion about methods of resolution Options

[Tman]

Hi Nirgal and all, I would like to discuss about vignetting and methods of resolution for it. I'm mainly interested in mathematical methods that could automatically calculate and adjust the right grey value for each pixel in a single picture.

My current method works with more or less transparent layers over the original picture that so roughly are able to balance the grey values. A perfect layer have to be the exact inverted brightness difference of each picture with this shadow effects. This method is very effective if you get the correct inverted values. These shots of the Mars sky come nearly at such a perfect mask, but not always. And of course the center of the pictures lose much of theirs original brightness/luminance sadly.

I have in mind a mathematical method that can adjust each grey value in a pic in order to obtain a completely balanced brightness over the entire picture. But I'm not in the position to reach that. I only know one have to start with the calculation of the grey values in the center of the picture. In the center are quasi the reference values of the whole picture, if I'm correct.
Is there a possibility (mathematical method) to get (roughly) the same brightness and luminance like in the center over the whole picture from the MERs?

Greetings, Peter

[MichaelT]

Hi Michael,

it would indeed be very nice if someone implemented it into MMB. Unfortunately, that won't be me as I do not have experience in Java-programming. I will post the code of my anti-vignetting routine in the next couple of days. So I hope that someone else can transform the code into Java.

Michael

[jaredGalen]

I would be interested in giving it a go. I made a java attempt at MichaelT's
DD enhancing algorithm before so I fancy a go at this even of some else beats me to it
But I think I might need a bit of expansion on parts of the algorithm posted at the start of the thread.

I'm not entirely sure of the different colour spaces.
For example can you convert directly from greyscale to LAB or must it be in RGB.
I have been looking and have found conversion from RGB to XYZ to LAB but not directly
from RGB to LAB.
>>0. convert the image to LAB color space to handle color and luminance independently

And then
>>6. so,as a last step I create a b-spline (or linear) interpolated version of the
correction curve.

In terms of getting the curve and things I'm a bit

I'll wait for MichaelT to post his routine I think, but I'll give it a go.

[MichaelT]

I would be interested in giving it a go. I made a java attempt at MichaelT's
DD enhancing algorithm before so I fancy a go at this even of some else beats me to it
But I think I might need a bit of expansion on parts of the algorithm posted at the start of the thread.

I'm not entirely sure of the different colour spaces.
For example can you convert directly from greyscale to LAB or must it be in RGB.
I have been looking and have found conversion from RGB to XYZ to LAB but not directly
from RGB to LAB.
>>0. convert the image to LAB color space to handle color and luminance independently

And then
>>6. so,as a last step I create a b-spline (or linear) interpolated version of the
correction curve.

In terms of getting the curve and things I'm a bit 

I'll wait for MichaelT to post his routine I think, but I'll give it a go.

Ok, its great that you want to try it I hope you find the following lines useful. By now I completely skipped using FFT . It can just as well be done without.

Concerning LAB etc: I don't know how to convert between the different color spaces. I found that the amount of vignetting in the color images is different between the RGB channels. So I don't know what it would be like if you applied the anti-vignetting to the two color-channels of LAB. Probably that gives some strange results? Don't know.

The other questions are hopefully answered below:

These lines of code are to generate the template function (see below).
First a vector T is generated that is XDIM long (XDIM: the x-dimension of the image) and runs from -pi/2 to +pi/2. The FINDGEN command outputs a vector of length XDIM containing values in ascending order from 0 to XDIM - 1.
EM = COS(T) * 0.1 + 0.9 is used as basis for the template matrix. As you can see it has a cosine shape. As initial value of the vignetting I assume 10%, ie. the factor 0.1. The 0.9 is added so that the maximum value is 1.0 while the minimum is 10% lower.
This vector is filled in a matrix of size (XDIM, XDIM) if XDIM > YDIM otherwise YDIM is used (IF statement).
To obtain a rotation-symmetric matrix, the transpose is added to that matrix and everything devided by two (ES).
To match the size of the image, the unneccessary data are removed (ESS).
To save time, the template is later fitted to a size-reduced image, and, therefore, reduced to the given size using the CONGRID-command. The size then is (512, YDIM*512/XDIM) or (XDIM*512/YDIM, 512) depending on the image (landscape or portrait).

[code begins]

IF XDIM GE YDIM THEN BEGIN
DISX = 512.0
DISY = YDIM * 512.0 / XDIM
T = (FINDGEN(XDIM) * 1.0 * !PI / XDIM -!PI / 2.0)
EM = COS(T) * 0.1 + 0.9
ES1 = {ES11: EM}
ES1 = REPLICATE(ES1, XDIM)
ES = (ES1.ES11 + TRANSPOSE(ES1.ES11)) / 2.0
ESS = ES[*, DIFF: DIFF + YDIM - 1]
ESS2 = CONGRID(ES[*, DIFF: DIFF + YDIM - 1], DISX, DISY)
ENDIF ELSE BEGIN
DISX = XDIM * 512.0 / YDIM
DISY = 512.0
T = (FINDGEN(YDIM) * 1.0 * !PI / YDIM -!PI / 2.0)
EM = COS(T) * 0.1 + 0.9
ES1 = {ES11: EM}
ES1 = REPLICATE(ES1, YDIM)
ES = (ES1.ES11 + TRANSPOSE(ES1.ES11)) / 2.0
ESS = ES[DIFF: DIFF + XDIM - 1, *]
ESS2 = CONGRID(ES[DIFF: DIFF + XDIM - 1, *], DISX, DISY)
ENDELSE

[code ends]

No comes the routine to fit the template to the image. It is rather complicated...
Using an iterative process I try to get as close as possible to the optimal fit. There are likely quicker/better ways to do this.

Firstly, ESS2, the template (see above) is now called E. The aim is to determin a parameter NR so that E^NR optimally fits the image.

First I calculate E^NR for three preset values given in INTR, yielding ESSR. Then I divde the image ( B ) by ESSR, which results in a temporary vignetting reduced image (FP_TEMP_R).
In step two I run through all rows (chosen by the user, actually rows between APY1 and APY2. That allows to do the fitting for parts of the image, eg. the sky, only) of FP_TEMP_R and fit a 2nd-order polynomial to the pixel-values (X is there to provide some x-values for the fit). The parameters of which are stored in RR. If an image is completely without vignetting, I assume that the sum PTR of the squared 2nd-order parameters (RR[2]) is minimal, as the "curvature" of the pixel values caused by the vignetting should be gone.

You might ask now, why I fit a 2nd-order polynomial to the supposed cosine shaped vignetting. Well, I found that the difference is not that large in most cases and that there are ready-to-use IDL routines for the polynomial fit.

As you can see I do this for the three values given in INTR [I1, I2, I3],. After doing so, I look for which of the two values PTR is minimal. I then use these two values from INTR (would be either [I1, I2] or [I2, I3])to define a new INTR with three new values. Two of which are I1 and I2 or I2 and I3 and a value in between the two (actually not quite I1, I2 or I2, I3 as the optimal value might still be in the respective other interval, but close to I2). The whole process then starts again with these new values. That is done a total of 15 times. After that the interval INTR has converged sufficiently close to the optimal value of NR.


[code begins]

E = ESS2

INTR = [0.0D, 15.0D, 30.0D]

X = DINDGEN(DISX) - DISX / 2.0D + 0.5D
RR = FLTARR(3, DISY)
RTR = FLTARR(3)


FOR K = 0, 14 DO BEGIN

FOR J = 0, 2 DO BEGIN

ESSR = E^INTR[J]

FP_TEMP_R = B / ESSR

FOR I = APY1, APY2 - 1, 1 DO BEGIN
WAIT, 0.00001
RR[*, I] = POLY_FIT(X, FP_TEMP_R[*, I], 2)
ENDFOR

RTR[J] = TOTAL(SQRT(RR[2, *]^2))

ENDFOR

SRTR = SORT(ABS(RTR))
INTR = [INTR[SRTR[0]], (INTR[SRTR[0]]+INTR[SRTR[1]]) / 2.0D, INTR[SRTR[1]]]
INTR = INTR[SORT(INTR)]
INTR = [INTR[0] - 0.05D * (INTR[1] - INTR[0]), INTR[1], INTR[2] + 0.05D * (INTR[2] - INTR[1])]

ENDFOR

[code ends]

All that can be done for all three color channels, as the amount of vignetting can be different for red, green and blue. Not for gray-scale images, though.

The code below shows what would be done for a gray-scale jpeg after NR (actually the parameter for the red-channel) has been determined:


[code begins]

NR = INTR[SRTR[0]]
NG = NR
NB = NR

ESST = ESS^NR

CR = CRS / ESST
CG = CGS / ESST
CB = CBS / ESST

[code ends]

Where CRS, CGS and CBS are the original color channels of the image. The new ones are CR, CG and CB.

In my program you can additionally choose howstrong the anti-vignetting is applied, so the final result would by something like:

CR = AV * CRS / ESST + (1 - AV) * CRS
etc., with AV between 0 and 1 (actually AV< 1 would be/is possible).

But I think this part is the most interesting for you jaredGalen. Most important is, I think, that the determination of the amount of vignetting can be done with a size reduced image. The found parameter can then be used for the original image. That saves a lot of time

The complete program can be found here

Unfortunately, I did not have time to comment it very much. I know, this is bad

Any other questions

Michael

[jaredGalen]

Wow, okay.

Looks like I have my work cut out for me.
Cripes, I'll have to pour over what you have said and your code first before I go
asking more questions. The color space issue is concerning me a little but I'll leave
that til I have a metter understanding of what you do.

I have a feeling I will have plenty of questions.

If anyone else outhere is thinking of doing this then by all means drop in.

Edit: Just realised the algorithm I was looking at originally wasn't posted by Michael. Muppetesque moment for me . That clears a few things up. Sorry about that.
Don't worry. I have everything under control