Real-time Document Registration with Deep Learning**

点击翻译按钮获取中文摘要

NeuroSoph is developing an intelligent form processing solution to classify and extract data from digitally scanned paper documents. A common issue during processing is the registration of scanned document pages to their respective templates. Registration is an image processing technique that aligns two pictures (documents in this case), so that if the contents of the two documents are overlaid after the alignment, they become congruent and indistinguishable. Registration of documents is a crucial step in the intelligent form processing. Scanned pages can suffer from various irregularities that must be accounted for in any registration process. These include the presence of user data, stains, creases, smudges, etc., as well as differences in color balance, scanner resolution, and scanner induced distortions. NeuroSoph is collaborating with University of Alberta researcher, Dr. Nilanjan Ray and his research team to find a solution to this document registration problem. The proposed project will find a robust solution for the rapid alignment of document images by employing modern tools of artificial intelligence, such as deep neural networks. This project will act as a technology pull for a Canadian startup firm, NeuroSoph, that seeks to develop and implement an intelligent document processing product and service. The project will also provide industry relevant training to two graduate students at the University of Alberta.