method: CITlab Argus Text Recognition2019-04-22

Authors: Tobias Grüning, Gundram Leifert, Jochen Zöllner, Tobias Strauß, Roger Labahn

Description: We use a deep neural network based on several convolutional layers stacked with bidirectional LSTM layers and a softmax output layer. The training loss is the classical CTC loss. The training data is consist of the given data, exclusively. We do not enhance the recognition by language models but return the best path only. To remove recognition results on accidentally found lines, we remove heuristically lines with no meaningful content.