Computing That Serves

End-to-end Trainable, Full Page, Handwriting Recognition

Curtis Wigington
MS Thesis Proposal

Friday, February 2, 12:00 PM
3350 TMCB
Advisor: Bill Barrett


Handwriting recognition (HWR) has seen significant improvements with the advancement of deep learning techniques. State-of-the-art methods usually require that a document first be segmented into handwriting lines before recognition. The majority of errors in HWR systems are a result of incorrect line segmentations. When line segmentation is treated as a preprocessing step, errors in the segmentation cannot be informed by the resulting errors in the recognition.
We propose a neural HWR system that perform both the line-level segmentation and recognition in an integrated, end-to-end trainable system. The system consists of three parts: the start of line finder, the line follower, and the handwriting recognizer. All three parts are pretrained on a small collection of documents with segmentation and transcription annotations. It then trains on a much larger collection of documents with only transcriptions.