How to evaluate image segmentation results?

I am doing with some fuzzy c means clustering based image segmentation extension work. Can please any one put the idea how to do performance analysis with some parameter with new segmentation approach.

Answer 1

[Image Analyst]

Have you tried tic and toc? Or the "Run and Time" tool on the "Home" tab?

[Tabish Raza]

yes !!comparison of execution for different algorithm can be done but me looking some other parameter that could best analyze the different segmentation approach.

[Image Analyst]

You Accepted, so are we done now? I'm not sure what the sentence after yes!! means. Are you saying that you want some other assessment of how well the algorithm works, other than the time it took?

[Tabish Raza]

i am looking for some evaluation parameter like PSNR,IEF etc which best describe your image segmentation evaluation.

[Image Analyst]

PSNR is not appropriate for determining how well a segmentation did it's job, neither is SSIM or any of those other image comparison or image quality metrics. Only you can decide if your algorithm got the parts of the image that you think it should have. For example you can use roipoly() and poly2mask() to create binary images that have the "absolutely correct" segmentation. Then see how many pixels match in your algorithm's answer and your hand drawn answer. Alternatively you can use some other algorithm that you trust, rather than hand drawing regions, and see how your new algorithm compares to the trusted algorithm. But either way you have to get the "true" binary image for each test image if you want to see how your new algorithm compares to the "gold standard, true" segmentation.

Answer 2

[Anand]

Two of the standard metrics used for image segmentation are dice overlap coefficient and jaccard index. These metrics measure the similarity between your segmentation and the expected segmentation output. This ofcourse means that you will need a "ground truth" segmentation result to compare against.

I found the following link that explains them nicely:

http://sve.loni.ucla.edu/instructions/metrics/

[Image Analyst]

Yes, those were the kinds of things I was thinking of. Nice to see that someone has thought it out more thoroughly. Thanks for the link.

Answer 3

[Sara Fadhil]

i need math-lab code or the syntax for dice similarity coefficient,variation of information,universal quality index,global consistency error,compare image boundary error,Davis bound,Jacquard index......any one can help for this

[Image Analyst]

See attached.

Answer 4

[Sara Fadhil]

I need Jacquard index to evaulate image segmentation algorithm.

I need Jaccard similarity code to evaulate image segmentation algorithm.

[Image Analyst]

From the help:

Description

example

similarity = jaccard(BW1,BW2) computes the intersection of binary images BW1 and BW2 divided by the union of BW1and BW2, also known as the Jaccard index. The images can be binary images, label images, or categorical images.

Introduced in R2017b

[Sara Fadhil]

thank you but matlab R2017didnt work on my computer

[Image Analyst]

Call them for free installation help if you can't launch your MATLAB release R2017 (or whatever version you have).