My project is Segmenting retinal vessels.
I use DRIVE dataset. In my project, I have one doubt. Can anyone clarify this.
[0 0 0 1 0 0 0; 0 0 0 1 0 0 0; 0 0 0 1 0 0 0; 0 0 0 1 0 0 0; 0 0 0 1 0 0 0; 0 0 0 1 0 0 0; 0 0 0 1 0 0 0 ]
This is one of the mask (T_d)(there are 12 masks like this). Green channel of a image is taken for the process.
T_di is centered on each pixel p of G and the following parameters are measured:
The arithmetic mean, Adi(p), of the gray-levels of pixels of G included by T_di and aligned along di.
The arithmetic mean, NAdi(p), of the gray-levels of all pixels of G included by T_di and non aligned along di.
The difference, diffi(p) = Adi(p)- NAdi(p)
Once all inspections have been accomplished, to assign the direc- tion to each pixel p of G the following parameters are computed, for k = 1,2, … , 12:
m(p) = mink {diffk(p)}
M(p) = maxk {diffk(p)}
This process is for identify the direction of a pixel. In this process we want to calculate these values for every pixel means how it is calculated. Please clarify it. Thank you.
Hi Kalai,
If you could possibly post an example image, I think our new software could really help you. We have tools specifically for measuring local directionality. At the very least it might help me better understand your problem and offer an algorithm solution.
Cheers,
John
Hi Kalai,
Thanks for posting! I am not sure if any of this will be helpful, but here are some example results of what our new software, MIPAR (http://MIPAR.us) can provide. I have attached outlined and B/W segmentations of the retina vessels. These were each accomplished using an automated 20-step recipe which involved a fine-window adaptive threshold, as well as morphology, clean-up, and masking steps.
From these segmentations, I calculated the orientation of the two-point correlation spot at each pixel, using a 10-pixel window (basically computing the average direction in a 10-pixel window around each pixel). I then colored each black pixel in the segmentations with their direction angle and attached these two direction maps.
If any of these capabilities would be helpful, feel free to sign-up for MIPAR at https://MIPAR.us/trial and I'd be happy to send you the recipe file! Let me know if you have any questions on the algorithm details and I'd also be happy to elaborate.
Cheers,
John
Hi Kalai,
if "drawing" your masks in 2 D you will get things like the following (generated from your mask values):
0 0 0 1 0 0 0
0 0 0 1 0 0 0
0 0 0 1 0 0 0
0 0 0 1 0 0 0
0 0 0 1 0 0 0
0 0 0 1 0 0 0
0 0 0 1 0 0 0
As you may see now, this is a pattern searching for a vertical line. The other 11 patterns will search for a horizontal line and maybe diagonal lines. Still too many matrizes I'd suspect you also have masks for wider lines :)
And the evaluation algorithm is looking for the best fit of one of the masks - giving you an estimate for the current orientation of a vessel crossing the current pixel. (As a "pixel" is 0 D, it cannot have an orientation. Thus I'd try to avoid to call it the "pixel orientation".
Hope this helps to understand "what's happening".
- Badges
- Science topic