How do I use svm to train and classify features extracted from a fingerprint image?

From the fingerprint image I have extracted three features which are a 1x16 vectors each. I don't understand how to use svm to perform the training and classification.

Answer 1

[Image Analyst]

SVM is mostly used for splitting clustered data into two classes, though it can be used for multi-class data but I hear it's a lot more complicated. How many classes to you have? How many classes are you wanting to end up with?

Note that the number of classes you want to end up with must be equal to or less than the number of data points that are used to decide upon which class something belongs. The number of classes is unrelated to the number of features for the most part. For example I could have an image with 16.7 million colors and a million pixels and 3 "features" (the red, green, and blue value of each pixel), and I could have two, a hundred, or a thousand color classes. Or as many as 16.7 million color classes if each pixel color was its own class.

[Tola]

I want to end up with two classes. I don't how to specify the number of classes.

[Image Analyst]

And what are those two classes? Why not use kmeans?

[Tola]

I am not really that good at using svm but from the little I understand. I want to create one class for the specific image features I want to recognize against all other features from all images I have. I don't really understand the concept of classes but that is what I think.

[Image Analyst]

No, sorry but that's not how classification works. And what do you think "all other features" might be? You can't do classification unless you know what they are. That doesn't make sense. The "other features" are, in fact, the same features. And you haven't defined the classes. Are your images all one class, for example whorl fingerprints? And then all your features define whorls? All right, fine, if that's the case, just threshold - real simple and no classification needed. You only need classification if thresholding doesn't work. Think of a 2D scatterplot. If you can't draw vertical and horizontal lines to box in your features, then you'll need to do classification. For example, let's say your image set consists of whorls, loop, and arches. Now if you plot a 3D scatterplot of your 3 features, you might see that the whorls cluster in one cluster, and the loops are in another cluster, and the arches are in a third cluster. Now if you can't just insert planes into that 3D scatterplot to divide them up, then you'll need to use something like kmeans. I think you'd better look for a tutorial on cluster analysis first before deciding on what to do.