Deep neural networks for record counting in historical handwritten documents

In this work, we study the use of Convolutional Neural Networks for counting the number of records in each page of an historical handwritten document. The initial network training is made on a large set of document images synthetically generated with a suitable tool implemented for this task. The trained network allows us to evaluate the number of records in the documents with a good accuracy that is subsequently improved with a fine-tuning performed with a limited number of real documents. In the experiments we compared three architectures on two datasets. On one benchmark dataset composed by marriage records we outperform previous results on a similar task.