Abstract by Jesse Williams
Handwriting recognition was once a very human task—computers have never been very good at complex pattern recognition in the way that humans are. However, in recent decades, advances in machine learning have facilitated the automation of handwriting recognition. Our research has been focused on reading historical documents, including US censuses. These documents present many unique challenges. For example, tabular data like census records are easier to read when segmented into table cells, which opens up possibilities for research about table segmentation. After cells are segmented, the handwriting in them must be read. Using transfer learning we hope to be able to apply the generalized deep learning model we create to various unrelated writing applications. The results of such research can be seen in apps like VetFinder.