Extracting Cracks

Hi,

I am working on an image which has pores and cracks. I was able to extract the pores excluding the cracks. Now I have to extract only the cracks. Is there any way to do this, that is to extract only cracks.

Below are the two images - 1st is the original image and second is the image which has only pores.

Below is the link for the image which has cracks rounded with brown color.

[Walter Roberson]

The second link is not found.

[Ratna]

Sorry for that, I changed the privacy settings, It should work good now.

https://picasaweb.google.com/117159985873958207679/November7201102

[Walter Roberson]

That looks like it is going to be a difficult task.

[Ratna]

I agree with you because the cracks are so faint. But I hope may be possible. Is there any way to do this or any command or anything like that?

[Ashish Uthama]

I can barely even see them at this resolution. What are these images? Can you improve the contrast/acquisition quality?

[Ratna]

@Ashish: These are CT images of plaster of size 512 x 512. I can't improve the acquisition because this is the best image we could get from CT scan. If I increase the contrast, I am getting unwanted/ unreal pores, making my image more blur.

Answer 1

[Sven]

Hi Ratna, It's quite difficult due to the relative intensities - cracks are only a "little" bit lighter than solid material. Here's my solution, which utilises Dirk Kroon's Frangi Vessel Filter on the FEX.

Sorry, I'm in a rush so I don't have time to comment it, but it's relatively short - hopefully you can follow line-by-line.

I've needed to put some pretty specific thresholds. I have no idea how it will work with "similar but different" images.

I = imread('Original_crackes.jpg');
Im = rgb2gray(I(130:700,450:1000,:)); % Clear away the border
Im_limited = double(min(max(200,Im),255));
Im_closed = imclose(Im_limited, ones(8));
Im_diff = Im_closed - Im_limited;
BW_holes = imopen(Im_limited<205,ones(4));
BW_pores = imclearborder(BW_holes);
Im_diff(imdilate(BW_holes,ones(10))) = 0;

% Get vessels
[Im_edgeEnhanced,Scale,Direction] = FrangiFilter2D(Im_diff/max(Im_diff(:)),struct('BlackWhite',false));
BW_cracks = bwareaopen(Scale==2, 50) & ~imdilate(BW_holes,ones(10));
CC = bwconncomp(BW_cracks);
stats = regionprops(CC, Im_edgeEnhanced,'Eccentricity','MaxIntensity','Orientation')
keepMask = [stats.Eccentricity]>0.983;
BW_cracks(cat(1,CC.PixelIdxList{~keepMask})) = false;

% Display
figure
subplot(1,3,1), imagesc(Im_limited), axis image
subplot(1,3,2), imagesc(BW_pores), axis image
subplot(1,3,3), imagesc(BW_cracks), axis image

[Ratna]

Actually my image is an .IMA image. If I run the same code using the Original_crackes.jpg, it works extracting one crack, when I use my original image - it shows an error at line 2.

The rgb2gray() function is used to clear the border, but in my image i don't need to clear the white border(as I don't have the that border in original image)

I am posting the error:

??? Error using ==> rgb2gray>parse_inputs at 82

MAP must be a m x 3 array.

Error in ==> rgb2gray at 35

X = parse_inputs(varargin{:});

[Walter Roberson]

The rgb2gray() is not what is clearing the boarder: it is the array indexing. If your image is already in grayscale and already does not have the boarder, then change the series of lines to something like

[I, cmap] = imread('Original_crackes.IMA');

Im = rgb2gray(ind2rgb(I, cmap));

I am guessing here that IMA images are pseudo-color images.

[Sven]

Yes, the initial setup (loading the image, converting to grayscale, etc) is tailored to the sample .jpg file that you uploaded.

Since your true image is from from a different source, you will need to adjust those first lines as Walter points out.

Keep in mind that some of the constants that I've used: - Image limits of 200 -> 255 - "Hole" threshold of 205 - Various sizes for opening/closing/dilating (specified using "ones(x)") These were also tailored specifically to the image I had at hand. If your true image is different (ie, it actually uses CT Houndsfield units for intensity rather than the 1-255 that a .jpg conversion gives), then you may need to adjust those constants.

If in your original image the large circle is covered by approximately 512-by-512 pixels (which is the size of most full CT images), then this matches quite well with the image size that I wrote the answer for, and the dilation sizes I used should remain about right. You're still welcome to tinker with them to see if it improves the output.

Answer 2

[Image Analyst]

Another option worth trying is anisotropic diffusion http://www.csse.uwa.edu.au/~pk/Research/MatlabFns/#anisodiff or coherence enhancing anisotropic diffusion filters, like is often used for fingerprints: http://www.vavlab.ee.boun.edu.tr/courses/574/materialx/PDEs%20in%20ImageProcessing/weickert_coherenceenhancing.pdf

Good luck.