IM Scientists Set New Benchmark for Weakly Supervised Text Recognition (OCR) at CVPR 2020
July 17, 2020
For the first time, IM scientists demonstrated a method ("OrigamiNet") to transcribe challenging handwritten text without line break info. OrigamiNet also achieves state-of-the-art OCR accuracy, even compared to fully supervised methods using line segmentation data.
Existing OCR techniques for handwritten text transcription typically require large quantities of fully annotated example data to train effectively. These annotations not only require the transcription of the characters themselves, but also manual segmentation of the locations of the line breaks.
For the first time, IM scientists have demonstrated a method ("OrigamiNet") that can transcribe entire paragraphs of challenging handwritten text, for example in old manuscripts or other hard-to-read documents, without having to supply this line break information. After applying the method to multiple benchmark datasets, OrigamiNet also achieves state-of-the-art results on character recognition accuracy, even compared to fully supervised methods that have line segmentation available.
The full paper on this method was published at CVPR 2020, the premier annual computer vision conference, and can be read here: https://arxiv.org/abs/2006.07491
If you are interested in learning how to apply this work to your own business problems, please contact us for more information.
