Offline Handwriting recognition using Bidirectional LSTM Neural Network
Handwritting recognition rates are low due to the difficulty in segmenting cursive or overlapping characters. Most progress in this field has been made in the pre-processing part, with little efforts in improving existing recognition algorithms.
This presentation talks about one type of recurrent neural networks, that is able to recognize data that is both hard to segment and contains long range bidirectional interdepencies.
The materials in this talk are from "A Novel Connectionist System for Unconstrained Handwriting Recognition" written by "Alex Graves, Marcus Liwicki, Santiago Fernandez, Roman Bertolami, Horst Bunke, Jurgen Schmidhuber" .