Code covered by the BSD License  

Highlights from
Anisotropic Diffusion (Perona & Malik)

4.36364

4.4 | 22 ratings Rate this file 147 Downloads (last 30 days) File Size: 8.01 KB File ID: #14995

Anisotropic Diffusion (Perona & Malik)

by

 

14 May 2007 (Updated )

A set of filters that perform 1D, 2D and 3D conventional anisotropic diffusion (gray scale data).

| Watch this File

File Information
Description

Anisotropic diffusion is a powerful image enhancer and restorer based on the PDE of heat transfer.

The implementation details are described in "P. Perona and J. Malik, Scale-Space and Edge Detection Using Anisotropic Diffusion, IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(7):629-639, July 1990" and in "G. Grieg, O. Kubler, R. Kikinis, and F. A. Jolesz, Nonlinear Anisotropic Filtering of MRI Data, IEEE Transactions on Medical Imaging, 11(2):221-232, June 1992".

The algorithm was implemented to perform upon 1D, 2D and 3D gray-scale signals.

Required Products Image Processing Toolbox
MATLAB release MATLAB 7.0.4 (R14SP2)
Tags for This File   Please login to tag files.
Please login to add a comment or rating.
Comments and Ratings (32)
13 Jun 2014 Stefan Karlsson

There is no productive 2D example provided that show the superiority of this method over regular gaussian smoothing. A simple combination of Gaussian and Median filtering provides better denoising than what is proposed:

s = phantom(512) + randn(512);
num_iter = 55;
delta_t = 1/7;
kappa = 30;
option = 2;
ad = anisodiff2D(s,num_iter,delta_t,kappa,option);

It is very likely that the authors generalization of using 8 point neighbourhoods is valid, but if so it changes the valid range of the parameters, and thus one cant easily recreate the results of the original Perona-Malik paper. If a simple example was provided that showed how the 2D method actually outperforms gaussian blurring, then I might be inclined to read the secondary paper referenced.

the interpretation of the kappa parameter is nicely given in weickerts book: http://www.lpi.tel.uva.es/muitic/pim/docus/anisotropic_diffusion.pdf, chapter 1, figure 1.1, b... (its called lambda there)

I cant get that interpretation to work for me here. Whats wrong? How can i strengthen the edges in the image using this, while blurring along isolines!?

Again, if only a positive example was provided for the 2D case, this might all be cleared up.

28 Apr 2014 SHANYUE  
19 Dec 2013 jia

very clear and have many details about the algorithm

08 Nov 2013 GIRE  
08 Oct 2013 Reena Singh Singh

I ve tried the code it is very useful bt i want to generate diffusion function to calculate flux function hw to do?

08 Feb 2013 Przem

the code here: http://www.mathworks.com/matlabcentral/fileexchange/31204-anisodiff-in-matlab/content/anisodiff.m

makes more sense. it is the actual 4 point PDE as propsed in the perona-malik paper

03 Dec 2012 Reena Singh Singh

thanx fr the code it is easy to understand bt in the perona-malik paper divergence is taken for the diffusion process nd in dis code diff-image is calculated in some different way. can u plz explain m dis?

07 Aug 2012 Andre

Arif Harmanci, if you input a column vector as required by the code you should get the expected behaviour. (in your example, just changing a_n to a_n' in the function call is sufficient).

11 Jul 2012 George Ntokas

I have a question:at the example in Anisodiff2D you wrote this line: s = phantom(512) + randn(512).The number 512 is about memory wright?But what exactly?And how i can use another image?Thank you in advance.

04 Jun 2012 Jiucheng Nie

for the 2D case: why all the flux from 8 directions are summed up? The paper suggests to use flus(north)-flus(south) and so on. There at least is 'minus' instead 'plus' all the way around. Could you explain me why?

25 May 2012 Arif Harmanci

I was looking at this code and I am pretty sure that there is a problem with the 1D code. I did not try 2d code. I tried following:

% Generate random regions.
a = [zeros(1, 2000) ones(1, 1000) zeros(1, 3000) 5 * ones(1, 4000) zeros(1, 6000) 3 * ones(1, 5000) zeros(1, 4000)];

% Add some noise
a_n = a + rand(1, length(a));

diff_sig = anisodiff1D(a_n, 1000, 1/3, 30, 2);

Then I get all 0's, that is, everything is smoothed.

If you go back to Kovesi's code, he did not use imfilter or conv, he computed the differences by basic vector subtractions. When I replace the imfilter to vector subtractions to compute nablaW/E, then I get smoothing only at the noisy regions, which is the expected result.

Would be great if anyone could comment on this.

25 Feb 2012 lina

Do you know how to change this function to be applicable to convert in C code using matlab coder??

12 Jul 2011 Tripp

Very useful and powerful filter

17 Jun 2011 Xianming Liu

Many thanks. That's just what I want.

18 Apr 2011 Faraz

Many thanks, but it gives me a blur image of a mamogram instead of enchancing the contrast?
can you/anyone comment y?

16 Mar 2011 Anand Muglikar

This is an awesome file that I used for driving out the noise in blood-slides and preserving the cell boundaries to detect malaria parasite.

06 Apr 2010 matlab user

thank you !

one question is that I can't duplicate the fig 3 in
G. Grieg, O. Kubler, R. Kikinis, and F. A. Jolesz, Nonlinear Anisotropic Filtering of MRI Data, IEEE Transactions on Medical Imaging, 11(2):221-232, June 1992".

the fig 3 show iterative edge sharpening, but my simulation gives me reverse result... can anyone give me a sample code for results similar to fig 3?

thank you in advance.

08 Mar 2010 Nitin

Thanks for sharing

04 Feb 2010 Joshua Groves

Thanks for this. I translated your code and created a Python version that I will hopefully be using for image analysis.

28 Nov 2009 Forrester Cole

Nice work. Though I think the examples could be a bit more helpful. For example, setting KAPPA = 0.2 or so, rather than KAPPA = 30, in the examples would show the anisotropic behavior more clearly. It took me a while to figure out that I needed to bring KAPPA way down from 30 before I saw results similar to Fig. 11 in the Perona and Malik paper.

28 Nov 2009 Forrester Cole  
07 Oct 2008 Solaree Shi

Thanks for it!

04 Aug 2008 Mohammad Ali Balafar

Thanks for your code. it is very clear and useful.

02 Jun 2008 Daniel San Martin Pascal Filho

Daniel meu chará, você me quebrou um galhão com essa implementação. Obrigado.

13 May 2008 Andrew Rein

Thanks for the code! it is very clearly written. For the 2D case in your code it is defined dx,dy and dd. What's the purpose of this? I didn't see it in the paper (have only read Perona). Also delta (lambda in the paper) suggest up to 1/4. Is it because you're using extra neighbours for the implementation?

09 Apr 2008 Daniel Mark

Thank you very much for this excellent code.

11 Feb 2008 albert j.  
02 Jan 2008 ivan scardanzan

Anisotropic Diffusion
see also on matlabcentral :
File Id: 3710 Average rating: 4.62
Size: 142 KB # of reviews: 26
Submitted: 2003-07-10 Downloads: 6020
Subscribers: 8
Keywords: NonLinear Diffusion, Anisotropic Diffusion, Coherence-Enhancing Diffus

20 Sep 2007 jianfei ge

Thanks!

20 Aug 2007 Bruno Carozza

Thanks, works perferctly!

22 May 2007 Daniel Lopes

Anisotropic diffusion is known to:
- smooth the signal;
- preserve strong edges;
- enhance the contrast of edges.

Contrary to spatially filtering the signal with a Gaussian, anisotropic diffusion doesn't smooth across the edges.

A recommend the following procedures to clarify your doubt relatively to the anisotropic implementation:

- run the example of anisodiff1D(.) [the visual effect of anisotropic filtering is more evident when applied to a 1D signal]

- instead of using a Gaussian with a small sigma try a higher one, e.g.,
hG = fspecial('gaussian',31,9); % sigma = 9;
Who will see that even if you increase the number of diffusion iterations the edges will prevail, contrary to the smoothing provided by the Gaussian filter.

Another advice: try reading and understanding the code with greater attention before rating it. The references to the articles aren´t there for nothing you now! :P

22 May 2007 David Gavilan

Is this really anisotropic? In file anisodiff2D.m there's an example using s = phantom(512) + randn(512); ad1 = anisodiff2D(s,15,1/7,30,2); Now try with a simple gaussian: hG = fspecial('gaussian',31,3); ad2 = imfilter(s,hG,'conv','symmetric'); Compare ad1 and ad2 and tell me if there's any difference.

Contact us