I have a series of images, all of which have been altered in the same way. A circle of radius, r, containing a face or other object has been edited out of the pictures and replaced with a best attempt at a background or blurred out entirely. The only information I've been given is the x,y coordinates of the pixel at the center of the edit, and the radius of the circle, r, though I've been advised that this might not be necessary. I need to create a neural network such that, given a photo edited in this way, can detect the x,y coordinates at the center of the edited disk. I think a CNN would do the trick? But whenever I do research on object classification or detection, it's always adjacent to what I'm looking for.

  • $\begingroup$ Why don't you simply use the difference of the edited and the regular image as an input? Then you probably do not even need a neural network for doing it because the only thing that will be unequal to 0 in the difference will be the pixels that have been edited and you just compute the middle point as if theyx were a square... $\endgroup$ – Fabian Werner Jun 2 '20 at 7:21
  • $\begingroup$ The dataset I was given is just the edited photos and the coordinates at the center of the edit. So when I'm given a new photo, all I'll be given is the edited photo with the need to produce the central pixel's X,Y coordinates. $\endgroup$ – Jayke Jun 2 '20 at 12:05
  • $\begingroup$ Ah ok... so why and how exactly is that what you find 'adjacent' to what you are looking for? Create some 2d convolution + max pooling + flatten and send that into a dense layer with two output nodes and relu activation (with pytorch for example)... $\endgroup$ – Fabian Werner Jun 2 '20 at 15:47

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.