Where to start assembling a neural network that places objects on a flat surface as efficiently as possible?

Question

Mathematics

Hello.
The thought came to mind, is there a neural network that can recognize geometric shapes, and then place them on the specified area as efficiently as possible, i.e. occupying for example the minimum space without imposing figures on each other. (for example, printing irregularly shaped stickers, the program will be effective for printing due to less paper consumption).
As I understand it, it is necessary to break the whole task into stages.
1. Recognize shapes by a closed contour from a source (for example, a scan of a sheet with a picture).
2. Calculate their area.
3. Start placing on the square.
How can you start collecting such a program?
matlab can it?
Maybe there are ready-made programs. Or pieces of code from which you can assemble the final version?

Answer 1

I agree with Daniil Smirnov that a neural network is not needed for this task. A neural network is far from a panacea for solving any algorithmic problems and will not fulfill any desire.
Matlab is suitable for mathematical modeling and testing hypotheses / ideas, but it is not suitable for the implementation of the target software product.
There is a wonderful library for image processing in the field of computer vision called OpenCV.
For the C# language, there is Emgu CV . For other languages, there are also bindings to the library. There was a good resource with examples, which has already died out, although it is available in the archive at the link opencv-code.com
In addition to what Daniil Smirnov wroteI will add that the task of optimizing the cutting of materials , which is one of the varieties of the knapsack problem and is solved by dynamic programming, is a well-known problem, even before the advent of computers.

Answer 2

So far, I have not quite found the use of a neural network in this task to
determine a sticker - that is, to recognize a closed contour - it is enough to filter the image with a sticker and select the contour using the same Hough transformation, and then determine the number of pixels (i.e. area) of this sticker with simple manipulations. do the same with other stickers. As a result, we have an array of areas of a certain shape - now we scatter them over the area you specified with another trivial algorithm