Assuming that one has a neural network capable of returning the numerical digit from a given image of size 28x28px.
How would one split an image of unknown size and an unknown amount of digits into a series of 28x28px images to feed to this network? (The order of the digits must be obtainable.)
How can this be split up to:
Assuming that there are not always 5 digits and the initial image is not always the same size.
Firstly, my thoughts would be to create a secondary neural network. This neural network would output an (x, y) coordinate. This could be used to crop the image to 28x28px about this coordinate. However, this neural network would only be able to locate 1 digit at a time.
Secondly, another idea is that a series of random crops could be performed and then all given to the number recognising neural network. However, this would create a high error rate and the number recognising neural network has no way of telling that there was no valid number given (Unless that output was added). But more importantly the order of the digits would/could be lost.
I am struggling to find any resource explaining a possible solution. Google's house number recognition takes the entire image of digits and returns a value. This can be found in How Google Cracked House Number Identification in Street View and Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks.