Overview / Usage

Recognition of offline handwritten text is a useful technique that can be applied to different tasks, such as digitalization of historical documents. However, the unconstrained handwritten text recognition is still considered a very challenging problem. The model presented is composed by an hybrid of convolutional and recurrent neural networks that recognize handwritten text using the image pixels as input features based on deep learning techniques. To train the model a Connectionist Temporal Classification layer was used, wich employs a cost function to align the sequence of label characters to the input sequence. A prototype of this model is implemented and trained using the IAM handwritten text database. The results obtained were satisfactory and show that the model is a valid choice for recognition of off-line handwritten text, beeing robust enough to be able to classify a pattern with so many variability such as handwritten text.