I am working on skin problem detection and am using a skin image dataset. I want to extract features from the images, but I can't understand which handcrafted features I should extract. I did some research on it. There are many features with LBP, HOG, HUE, and so on. There are numerous. Do I need to remove and concatenate those before passing it to an algorithm? I mean, if I'm exacting HOG and LBP, will I merge them both and then give them on to the ML or DL algorithm?
Add a comment |
1 Answer
$\begingroup$ $\endgroup$
1 one approach could be to process the images with classic CNNs and the metadata/additional features separately, for example with a Fully Connected NN. The outputs from the two networks can then be concatenated and fed into an additional FC NN to generate the final predictions.
Check this repo for the source code and further implementation details!
- $\begingroup$ actually, I wanted to extract handcrafted features from the images, and merge that with the features extracted from CNN. But I think the repo you shared concatenates the image data with tabular data. $\endgroup$Rezuana Haque– Rezuana Haque2022-12-04 12:38:03 +00:00Commented Dec 4, 2022 at 12:38