0
$\begingroup$

I want to solve OCC(one class classification) on images, the input would be an image, the output is if the image belong to the class, and I extract the image feature from a deep neural work, the feature is in a very high dimension, more than 25000, I find that the Nearest Neighborhoods would be less meaningful when the point in a high dimensional space on this paper, the LOF(Local Outlier Factor) algorithm wouldn't make sense.

Can someone point me in right direction?

$\endgroup$
1

1 Answer 1

1
$\begingroup$

Your two best choices are either dimensionality reduction or another algorithms that doesn't rely on distance measures.

If you have plenty of images, you may train a neural network and don't need to do dimensionality reduction or other feature engineering.

If you don't have so many images to train a neural net, try to find a way to extract features from your images first and classify on those features.

$\endgroup$
1
  • $\begingroup$ Thanks for your advice, but it seems that I'm facing an OCC problem, I will never have enough images on the negative class, It's easy to find images showing a specific object, but hard for images not showing the object. And I don't have too much knowledge on image processing, I'm afraid I can't get enough features from the image, that's the reason why I extract features from pre-trained convolutional neural networks. I wonder if I can solve the problem based on this paper. $\endgroup$
    – Craig.Li
    Commented Sep 26, 2017 at 13:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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