If a use a feature extraction method on images, do I then get a feature representation or is there a different meaning behind feature representation?

To my understanding, when I use a CNN on an image the convolutional layers perform feature extraction. Can I then also say that I get a new feature representation after each of these layers?

Thanks in advance! I am relatively new to machine learning...


1 Answer 1


Yes, you are right. As a result of the extraction of features, new feature representations are created. For example, SIFT and HARR are algorithms to extract the crucial features from the images and create feature representation. CNN does it automatically with the help of convolution layers. In an image, smooth area, edge, corners are considered features that may represent it best. The importance of the smooth area, edge, and corners (there are others too) are normally accepted in ascending order.

In a simple CNN, if you can visualize the layers you can see that what it extracts from the image are mainly, corners and edge. It is also able to learn more complex features than we know (or defined). So, as you said, the convolution layers try to create a representation of features by extracting them. Then, this features representations (important information from the image) are used in the let's say prediction process.

  • $\begingroup$ Thanks a lot! With your explanation I understand it more now :) Are these visualized layers then extracted features and a feature representation at the same time? $\endgroup$
    – mar_ey
    Nov 7, 2020 at 12:55
  • $\begingroup$ Yes. To "extract feature" it does not have to be an original image, since CNN normally has many convolution layers, you extract the feature from feature representation which this feature representation was an extracted feature in the previous layer :)). I hope it is clear. $\endgroup$ Nov 7, 2020 at 13:02

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.