The choice of architecture and means of implementing a convolutional neural network?

Question

Python

It is necessary to determine the architecture of the convolutional NN, as well as the means for its implementation for the following task:
NN is needed for a visual quality control system for products (detection of defective caps).
The input is 1000x1000px images (You cannot compress them, as small defects may disappear).
In the work will need to classify 60 images per second! (Is it possible?)
Calculations are done on the GPU (GTX-1080 Ti)
If I understand correctly, this is a classification task. The National Assembly during training should highlight the signs of a defective product and determine: marriage / not marriage.

• What architecture can be used to solve this problem?
  
• Is it possible to use existing architectures (AlexNet, VGG, ...) and optimize them specifically for this task?
  
• Is it even possible to use a convolutional neural network for this task? Will it cope with the detection of marriage (for example: small black dots or an irregular geometric shape)?
  
• What is the best way to implement such an NS?
  
• What is the best framework to use?
  
• What programming language: Python/C++/MATLAB?
  

I beg you to help competent people in this area!!!
What can be read on this topic and how to deal with it ??

Answer 1

In fact, the answer to the first three questions asked is not sickly such a job. With the corresponding analysis, comparison, etc. necessary attributes.
The answer to the last three is at the level of a freshman-three-year student (excellent students, even in the first year, don’t ask such questions anymore).
So based on the questions asked, you have two options. The best thing you can do - if you do not want, of course, to fill up the project with a bang - is to hire a person competent in these matters. Expensive, of course, but you have to pay for the knowledge and skills (of others).
The second way is to independently understand the topic, the benefit of literature is in bulk, it’s even somehow not very convenient to list them here. Well, for starters, except perhaps:
1. Haikin Neural networks full course. 2 edition
2. Nikolenko S., Kadurin A., Arkhangelskaya E. Deep learning. - St. Petersburg: Peter, 2018
3. Goodfellow Ya., Bengio I., Courvill A. Deep learning - M .: DMK Press, 2018. 4.
David Kriesel A Brief Introduction to Neural Networks
then they offered to master the doctoral dissertation along the way :-). True, I don’t think that you will understand it without mastering the origins)
Well, there are more than a lot of links on the Internet. For the very lazy - there are video courses. Even in Russian :-)
Only this long way. So the first one is better.

Answer 2

Go to ODs.ai, join a group and ask the experts there. At the same time you can find a performer there. This is a machine learning slack group and it has people who have won image classification competitions.

Answer 3

https://github.com/RedditSota/state-of-the-art-res...
Here are links to SOTA (state-of-the-art) architectures for various tasks + their implementation in various languages. You can take some solution with results on a more or less normal dataset (not on MNIST), figure out how to put the data there and scroll.