5
$\begingroup$

I have an image. I find features on it using different methods. I want to quantify that which method performs the best. The creteria is how many features per unit area of the image are and how they are spread. More the features are spread, the better it is. How to visualize them.

So at the end the criteria is "how many features per unit area are found and how well are they spread all over the image"

It will be great if there are some suggestions that how to visualize the quantified results.

Thanks a lot.

$\endgroup$
2
$\begingroup$

To visualize the features, why don't you just do scatter plots? I assume that if you do feature extraction, you know where in the image the features are found. If different features have different "sizes", i.e. some features are point features, whereas others take up e.g. a 5x5 window (and are non-overlapping), you may want to adjust the size of your symbols for the scatter plot to visually improve the relative amount of coverage.

If you want to quantify the coverage, then you're looking at two independent measures

  1. Density: To this end, you can simply divide the number of features by the image size.
  2. Homogeneity: I've seen Ripley's K used to great success when looking at whether data is homogeneously distributed, or whether it is clustered.
| improve this answer | |
$\endgroup$
1
$\begingroup$

The hardcore approach is to measure the mutual information between your input and your features. I have once applied this approach to audio with some degree of success. Take a look at this http://www.bme.ogi.edu/~lantian/bibo/feature%20selection/Input%20Feature%20Selection%20for%20Classification%20Problems.pdf

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.