Deep learning/Neural networks/Machine learning

I own a system developed for a very specific industry in photo sorting. One of the most overheading tasks currently done by human employees is to compare two photos and say if the photos are repeated or not. For example, when a photographer takes two consecutive shots of the same group of people, the photos are usually repeated. However, there is a very subjetive criteria on when the photos are repeated. I would like to automate this task. Check the attached photos to see an example of repeated/non repeated example that the system is supposed to classify. Also, the employees have to choose which photo is the best one when there are repeats. I need a system to do that too. For the training, I currently have a database of 50 thousand categorized photos, and if you prove your system works, I can categorize 1 million photos in one month for it to train with. I don't care about which software, language, algorithm or API you use, if you deliver 99,5% precision compared to human performance. 99,5% is the minimum, and anything worse than that is useless. You are welcome to use Azure's machine learning platform if you find fit, or any other platform that doesn't exceed the maximum of US$0,001 per photo.