I have two kinds of features (say "intact" and "cracked" samples) in my image data. How do I classify new set of samples into "intact" or "cracked" samples based on my already known samples for training and using the image parameters of the training set?
Hi, I think that il not easy to see with nude eyes the difference between "intact" and "cracked" images. Do you have defined new parameters (not based on pixels colors, or contrast or any other classical params) witch depends on image meta-information.
By defining this parameters, I think that it is possible to use an ANN or Naive bayes methods to recognision.
I hope that I'm not so far from the goal
Best regards.
Thank you. Your answer will push me further. Actually my challenge is in defining such image parameters that might be useful for such recognition. Im still very new in this area of image processing.
What you mean by 'features' is usually referred to as labels or classes in literature. What image parameters do you have? if they are nice (not too high dimensional vectors) you might want to use WEKA with your training set, build a classifier and then run it for the unseen set (new set of images) to see how they are classified. I think this would be the easiest way and you can try several different classifiers to see how they behave on your image parameters.
As Juan correctly noted, what you mean by 'features' is usually referred to as 'classes'. In fact, what you need as a first step is to extract meaningful 'features' from your input images using a proper image descriptor. It is difficult to foretell which could work well in your application without having a look at your images. Could you post some?
Look into classification methods, that are generally part of the machine learning field. The main idea is that through presenting the training set, your system will learn to extract some protypes, or classes, in your case two classes "cracked" and "intact".
So you present an image to the network and you tell him "this is cracked". And you repeat that for the whole training set.
Then when you present an image, the network will output which class is associated to this image.
You can look at a tutorials on perceptron and multiple layer perceptron for a very fundamental understanding about the computations done by classifiers (i.e transforming input space to an hidden space that is linearly separable).
Good luck for your personal learning curve ;)
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification