< Back to previous page
Project
IOF PoC “CAPTCHA 2.0 - Completely Automated Processing of scanned Text documents by teaching Computers how Humans Analyze them” (IOFPOC40)
The POC project CAPTCHA 2.0 - Completely Automated Processing of scanned Text documents by teaching Computers how Humans Analyze them builds further upon the well-known CAPTCHA's.
These were introduced to prevent spam attacks by exploiting that computer vision has not (yet?) reached human capabilities, meanwhile helping with image-to-text processing of old book scans. Humans are indeed superior in recognizing text from complicated layouts, be it because of typesetting or quality of a scan. We can read with ease glossy magazines, bilingual articles or old manuscripts where letters are distorted because of creases or tears, by combining visual cues with layout and language understanding. Although Optical Character Recognition (OCR) is around since the 90s, the performance of existing text recognition software is still too low for companies to use it in a fully automatic document processing pipeline, mainly because of the lack of a satisfying layout understanding. This is exactly where our PoC comes into the picture (pun intented).
DIMA developed an AI technique - Document Segmentation with Probabilistic Homogeneity (DSPH) - that mimics the human visual processing of document images which, when applied right before OCR, can bring automatic processing of scanned document images to the next level. This has a huge potential for document workflows in industry (e.g. automatic processing of scanned invoices, payslips ...) where human interventions/interpretations will be drastically reduced or even removed.
The project aims to bring the technology to the market.
These were introduced to prevent spam attacks by exploiting that computer vision has not (yet?) reached human capabilities, meanwhile helping with image-to-text processing of old book scans. Humans are indeed superior in recognizing text from complicated layouts, be it because of typesetting or quality of a scan. We can read with ease glossy magazines, bilingual articles or old manuscripts where letters are distorted because of creases or tears, by combining visual cues with layout and language understanding. Although Optical Character Recognition (OCR) is around since the 90s, the performance of existing text recognition software is still too low for companies to use it in a fully automatic document processing pipeline, mainly because of the lack of a satisfying layout understanding. This is exactly where our PoC comes into the picture (pun intented).
DIMA developed an AI technique - Document Segmentation with Probabilistic Homogeneity (DSPH) - that mimics the human visual processing of document images which, when applied right before OCR, can bring automatic processing of scanned document images to the next level. This has a huge potential for document workflows in industry (e.g. automatic processing of scanned invoices, payslips ...) where human interventions/interpretations will be drastically reduced or even removed.
The project aims to bring the technology to the market.
Date:1 Jan 2021 → 31 Aug 2022
Keywords:document analysis, segmentation, OCR
Disciplines:Computer science
Project type:Collaboration project