You can try to learn NN with examples that are also easy to distinguish for your eyes. Try it with different colours, different shapes,...
If you already have the collection of images then you can search for patterns that have the nearest neighbour of different class most far away (manhattan or euclidean distance).
This type of problem is known as Landmark recognition problems . Commonly many people use pre trained models such as VGG, RestNet Inception and Yolo rather than training from scratch.
1. Download the pretrained models of any of these models from internet , it is free.
2. You need to have a file which annotates the markers. You can try Yolo, which outputs a confidence score and the coordinates of the bounding boxes. You can read the Yolo paper.