Fast median algorithm

[Siekmanski]

[guga]

This is what shows to me

[Siekmanski]

That's not okay.
Work to do.... 

[guga]

 


This is with the version i did with the normalization. (This one needs a small fix. It is the  outdated version i used to compute the Sobel alone and not insid ethe main routine of the watermark remover. I´ll later join them both)

[Siekmanski]

That's how it supposed to be. 
Your post of my distorted image gave me a hint, it probably is the routine that draws it to screen.
I'll have a look tomorrow.

[guga]

Hi Marinus

Finished analyzing the Laplace algorithm and this Lambert W Function. I succeeded to create this W function (with SSE2  )  and i´ll give a try to see if it is worthfull to use laplace as an edge detector. I also implemented a log and exp function using SSE2, but the code is a kinda mess. So i´ll pout them later (here on on another thread to we test the speed :) )

Attached is a tutorial i wrote about this Laplace stuff to don´t forget what i´m doing

[guga]

Hi Marinus


I-´m not sure if i implemented this correctly, but...if i did, then....this is what Laplace results

[Siekmanski]

Hi Gustavo,

Great. 
I will digest the information you wrote, later this week.
Have a busy week ahead, so less time for coding.

[guga]

No problem 

I´m still trying to understand this whole laplace concept. The result using is 2-D Log function with Gaussian (without those matrices) is a sort of a mask. So i don´t know if i should use these 256 new values as a 16x16 matrix (that will be biased on each image) or if i should try to search the pixel values and compare with the results of the table (This is what i did, in fact).

I´m watching these videos
https://www.youtube.com/watch?v=kJKgCwUzkmc
https://www.youtube.com/watch?v=zQKNVept4bU

to try to understand.

Well.. at leats i got that damn W lambert function working with SSE. Despite the results i´ve got, the algo is pretty fast

[Siekmanski]

Cool.

So, you're in love now with SSE... 
Maybe it is possible to transfer it into a matrix with pre-calculated coefficients???

[Siekmanski]

Just watched the first video, it shows the coefficients for the 3*3 matrix.
As I understand it, it's used as a filter.
If so, it has to be used in a step before you apply the Sobel operations.

[guga]

Yes, i´m liking to learn SSE

A bit painful at 1st, but the results are great.

What i did not understood is what a hell is this thing doing ?



I succeeded to create a formula and also made a function to work with this W Lambert stuff. But, what i did was assume that Log(x,y) = 0 to 255 and i´m not sure if it´s correct to do this way.  The resultant image is a sort of a mask but i´m not sure if all i did was correct. I tried to not use the matrix values and go straight to this formula.

The result was 256 values i used as a table, but i have no idea if it´s correct to do that way.

What´s log(x,y) all about ? I know what is log(x*y), log(x/y) etc, but with a ",", i have no idea.


And also, the above formula is different from the one below:
https://academic.mu.edu/phys/matthysd/web226/Lab02.htm



Here it uses a power of 4 to calculate the Standard deviation, but on the image it displays a power of 2 

[guga]

Ok, i guess i found how this stuff really works.

It´s used a a kernel generator, in fact

https://www.geeksforgeeks.org/gaussian-filter-generation-c

[daydreamer]

Cool.

So, you're in love now with SSE... 
Maybe it is possible to transfer it into a matrix with pre-calculated coefficients???

learn from RCP** instructions the backward math:way of faster coding together with replace x DIV** coefficients to x MUL** 1/coeffcients with upto REAL8 precisicion

[Siekmanski]

Thinking it over, I think we can skip the overcomplicated approach.
We are not trying to get the highest accuracy.
Not an issue with only 3 coefficients.
Let me try to explain what I mean.

Let's think backwards:

The whole idea is to find coefficients for a lowpass filter.
The choices are, how do we want the curve to behave.
In other words, what type of filter.

filter example in 1 dimension:  [ -0.5 1.0 -0.5 ]

For the computation of a 3*3 matrix we only need 3 main coefficients, which represents a 2D first second order filter.

The Laplace filter uses the Standard Deviation to calculate the curve properties.

For 2D Laplace filter, it turns out to be these coefficients:
Very effective for the edge detecting routines. ( you see the cross in it? )

in 2 dimensions:

[ 0  1  0 ]
[ 1 -4  1 ]
[ 0  1  0 ]

1+1+1+1-4= 0

or try something like this:

[ 1  1  1 ]
[ 1 -8  1 ]
[ 1  1  1 ]

1+1+1+1+1+1+1+1-8= 0

With so few coefficients you could create your own filter type, by experimenting with the values.
As long as the surrounding values add up to zero including the middle value, you have created a filter with your own roll-off factor.

This could be a nice one too:

[ 0.5  1  0.5 ]
[ 1   -6  1    ]
[ 0.5  1  0.5 ]