robertknight
Feature request: Add confidence scores to library output(s)
It would be really nice to have some sort of (maybe even optional) confidence scores for text recognition output.
It seems that information is already present in the lower-level libraries so it would be just a matter of exposing and/or normalizing it for the library user.
For the default DecodeMethod::Greedy decoding of the neural network outputs there is a score per character that could be exposed. If the beam search decoding method is used, it gets a bit more complex because there is a score per line which is a sum over multiple "paths" through the outputs. The raw scores the network computes are log probabilities (output of a LogSoftmax operation). It might be more user friendly to convert them back to probabilities first.
