How to organize a neural network to solve the problem of object localization?

Question

Neural networks

Hello, members of the Toster project.
Could you help clarify some details?
The essence of the question/problem is contained in the title.
Whether the correct approach is used:
An image is fed into the convolutional neural network, there are several layers through which the input data passes.
The output is n neurons, each of which is responsible for a certain area of ​​the image.
Thus, I would like to get a mask that contains m of n areas from which the "signal" was the strongest (above the threshold).
Input data: images and masks with which the output of the neural network is compared.
So, as planned, in comparing the mask and the output, we are looking for an error function.
Is the right approach chosen?
If not, then what configuration and approach would you recommend?
Or is it impossible to solve this problem only with the help of CNN?
Ps The decision time is not important.

Answer 1

I'd rather look at https://pjreddie.com/darknet/yolo/ or https://github.com/yhenon/keras-frcnn

Answer 2

Google the phrase: u-net neural network
I think this model is just what you need, and it has proven itself in medical image segmentation, and in other machine learning tasks (for example, Kaggle Sea Lion Population: Estimating the Sea Lion Population... )