I recently learned about face recognition with deep learning here. One of the approach involved is Histogram of oriented gradients which is used for face detection as follows (short summary) :

  1. convert image to gray scale
  2. look at every pixel in image and detect surrounding pixel
  3. draw and arrow in direction where surrounding pixels are getting darker
  4. repeat the whole process

this results in HOG version of image ,something like this :


Question Can we use similar method in colored images like how does rgb varies at pixel level or how does pixels vary in colored images to improve the accuracy of face detection and similarity.

Purpose face comparison is a bit tedious in this approach in terms of accuracy. So, i want to find some sort of hash function/value(just using as layman term) to derive unique value from every image. This might improve face comparison easier.

Additionally , sharing any already implemented approach in this direction will be highly appreciated.

  • $\begingroup$ i believe the questions doesn't typically match SO standards, but i needed some advice regarding it $\endgroup$ – parth Aug 3 '17 at 9:10
  • $\begingroup$ If you can narrow the question down to something that shows some effort, and is both clear and answerable, then it should be fine. Are you asking whether the accuracy can be improved by using the three color channels instead of converting the image to a single value channel? I also don't understand the purpose paragraph, can you rephrase it? The purpose of the embeddings are to obtain descriptive information to each captured face. $\endgroup$ – E_net4 is cleaning up Aug 3 '17 at 18:34
  • $\begingroup$ appreciate your words @E_net4, i'll rephrase the question soon and yes, i want to know whether accuracy can be improved using multi color channels instead of single value channel ? $\endgroup$ – parth Aug 4 '17 at 5:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.