How to competently approach the recognition of objects from images?

Question

Neural networks

I got a task to recognize objects from satellite Google maps. By any means, to achieve at least that the building was in the picture.
The difficulty for me is that I don't know where to start. So far, I started to study Opencv in Java (I have been writing in this language for 2 years). Questions immediately arose about how to search for the same building in the picture - look for some ready-made library tools or saw your own neural network. From here, new questions immediately arise:
- what should be the neural network (before that I wrote a simple perceptron for digit recognition), how many layers, how many inputs (here the question follows)
- whether and how to split the image into separate sections of pixels by 100x100 or 50x50 and etc.? because it depends on the number of network inputs
- what morphological transformations of the image should be performed using opencv tools?
- Is a neural network really needed here?
I will be glad if there is a person who will at least show in which direction to dig, and give a couple of tips.

Answer 1

Dig towards "visual pattern recognition". Perhaps one of the few areas where a neural network is the most suitable means of solving a problem. Brief algorithm: you take thousands of images on which there are and about the same number - some of which do not have objects of the desired class (buildings), train the neural network, get a network that is capable of with some probability (depends on many factors - the quality of the training sample, the selected network class, including the number of layers and types of connections between them, the correctness of the learning process, etc.) to recognize the building from the image. You have to start with theory. Most books on neural networks have a description - or at least a detailed roadmap - of solving this class of problems. The tool (both software and platform) is selected based on the capabilities of the customer.
Good luck.