Ultimately you'd be training a net to identify linear features arranged in consistent pattern. The feasibility of such a net depends on the signal to noise rati

Ultimately you'd be training a net to identify linear features arranged in consistent pattern. The feasibility of such a net depends on the signal to noise ratio in the input data (how subtle are the encoded words?), with the number of images required to train to a minimum accuracy increasing with busier input data. This assumes you're training a net with a suitablly complex and appropriately architected neural network. For some degree of intuition regarding training data, I like to imagine neural networks as infant brains which will be shown classified training images repeatedly in order to teach associations. Unfortunately we haven't really come far enough yet to give concrete numbers. tldr: might work, if you have a suitable architecture, and enough training data, in proportion to the complexity of the encoding.