Handwritten text recognition in historical documents

Thesis with the following contributions:

• Analysis of different neural network architectures and parameters
• Word segmentation using the output of the RNN layers
• CNN-based replacement of the RNN layers (enabling a purely convolutional architecture)
• Constrained CTC decoding algorithm (see paper for more details)

Word Beam Search: A Connectionist Temporal Classification Decoding Algorithm

Paper presented at the 16th International Conference on Frontiers in Handwriting Recognition, 2018, Niagara Falls, USA. Properties of proposed algorithm:

• Decodes output of CTC-trained neural network
• Words constrained by dictionary
• Allows arbitrary number of non-word characters between words
• Optional word-level language model
• Faster than token passing

Comparison of Connectionist Temporal Classification Decoding Algorithms

Comparison between beam search decoding and token passing. The paper was written as part of a course at university and was not published at a conference. Nevertheless, I'm presenting it here because it offers some practical tips on using those two algorithms.