Handwritten Word Recognition on the Fundación-Osborne Dataset

Even to this day, offline handwritten text recognition still constitutes a challenging research problem, specially when it comes to perform recognition tasks on historical databases. In this context, the main aim of the present paper is to expound the results obtained after training a deep convolutional Seq2Seq network with attention mechanism using a combination of word training images from both contemporary and historical databases. In the light of the subsequent results, we discuss the effectiveness that different proportions of modern and historical text during the training process have on the final performance of the architecture concerning historical handwritten text recognition.