Authors: J.Ignacio Toledo, Manuel Carbonell, Alicia Fornés, Josep Lladós

Description: The sequence tagging is performed with a CNN+BLSTM neural network, that accepts a sequence of word images as inputs and produces two independent softmax outputs (person and category) for each word. Those word images tagged as meaningful are then transcribed by generating a sequence of PHOC embeddings that are fed into a BLSTM+CTC network.