Publications
Handwritten text recognition in historical documents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