1
$\begingroup$

I aim to train my classifier for an image recognition task. What kind of preprocessing steps I need to take to enhance my results? (I have >40000 images with >700 pixels so large amounts of data with a vast feature vector.)

Also can you please suggest any other efficient preprocessing method to advance performance after you share about normalization, standardization?

$\endgroup$
2
$\begingroup$

The need for feature normalization depends on the model you are using. Some are resistent to the shape of your features, others are strongly connected. Here are some properties you may want to have :

  • a fixed size for each image. Do not hesitate to work first on small sized image (like 64x64px or even 32x32)
  • keep only the grey level for all pixels (dropping color reduce the number of features by 3).
  • normalize the contrast of your images
  • try to work on a gradient map instead of the image itself
  • other factors may interfere, like the orientation of the objects, the shape of the background, etc... A well chosen kernel can deal efficiently with them.
$\endgroup$

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.