How can I implement an analog of msqrd in an image web application?

Question

Programming

Greetings colleagues. There is a task, while I don’t even know which way to dig, perhaps even the term computer vision does not fit here. Can you tell me ready-made tools for implementation, or at least keywords for which you can search)
TK is like that. The client uploads a photo where, for example, a 5-pointed star is captured on a contrasting background, or a person, let's say, in full growth. Our script defines the contours of the object, as well as the fill area, let's say everything except the head, hands and feet. Well, let's say the belt divides into 2 areas, conditionally pants and a jacket. Next, we have 10 types of texture, clicking on which fills these areas. Roughly speaking, analog of a computer. games where we dressed up our character, only here it’s not a 3d model and the 2d model is taken from a photo. The fact that it will be a texture is not important, adjusting it so that it is realistic is zaparnenko, but for now I expect to implement the idea in a couple of days or a couple of hours)) Filling with a texture as a pattern is already cool.
How to fill in an area is basically clear, js, for example, through canvas or something like that. But how you can "parse" a photo for such areas is a mystery.

Answer 1

Too little information, but in any case, a couple of days - well, a VERY optimistic estimate. I would increase it to at least a couple of months only for a primitive prototype.
as an option (perhaps not quite correct and optimal, but working):
1. definition of contours (prewitt / Sobel methods)
2. classification of the contour (to understand what we found, and what options for "replacement" can then be used). I would look here in the direction of perceptrons / convolutional networks (yes, they still need to be trained)
by itself, you will need to process the image on the server (I would suggest a bunch of python + theano / tensorflow / skilean) I have never worked with OpenCV, but it is possible that some of the required functionality has already been implemented in it (at least there should definitely be something about highlighting contours).
after it comes out to train the neural network on the serverside, it will be possible to try to transfer the entire solution to the browser side (outline selection algorithms + neural network + coefficients that we received after training it on the serverside).