3
$\begingroup$

I am working on image classification on a really small dataset which contains around 220 images with two labels (110 images for each label). My goal is to visualize the common feature extracted among each of the classes/labels to understand why certain images are classified as the first label instead of the second one. The images are labeled by experiment and we are trying to understand why they are labeled in that way by visualizing the feature.

What would be a good approach for this problem? I've considered deep learning but I think the dataset might be too small for it. And also what would be a proper practice to visualize the features?

$\endgroup$
  • $\begingroup$ Please register &/or merge your accounts (you can find information on how to do this in the My Account section of our help center), then you will be able to edit & comment on your own question. $\endgroup$ – gung Sep 5 '17 at 1:37
2
$\begingroup$

Two approaches come to my mind: using using some Fourier-like transform for extracting features or pretrained NN model.


Fourier and wavelet transforms can be thought as representing image in basis that captures frequency components or something between spatial/frequency components (in uncertainty principle sense).

For an example of usefulness of this approach see An Introduction to Compressive Sampling. For some concrete examples see these notebooks (this project uses PyWT).


Using pretrained models is another option - there are pretrained weights for images used for example for classification tasks. Using something like this is also called transfer learning. It basically consists of dropping the classification layer and using the network as encoder, or maybe then appending another classification layer.

I'm no expert on that, but from Keras documentation for example, it seems that it's pretty straightforward.

$\endgroup$

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.