16 March 2017: Tested a Trial Copy of ABBYY FineReader

Experiments

• In order to better judge what's possible for OCR, we are sampling both proprietary and open-source softwares
• We installed ABBYY FineReader 12.1.x onto the dorkroom mac mini
• Asked it to convert the images from the previous experiment with Tesseract
• It produced a PDF containing the first three images (File:Test001.pdf) -a limit of their trial version- with the following issues-
  • Positive
    • Pages were automatically oriented for English LRTB
    • Pages were automatically straightened
    • It produced indexed, searchable PDF
    • It indexes scientific terminology, including possibly unrecognized words - a bonus for scientific literature
    • It recognized images, tables and diagrams, and paginated them to match original layout in the resulting PDF
  • Neutral
    • The resulting PDF contains not only text, diagrams and images, it also contains the entire original scanner image
  • Negative
    • None of the pages was automatically cropped, so the scanner platen occupies most of the image
    • One of the pages was cropped, by FineReader, but incorrectly (page 3), removing all text content
    • FineReader is not fast.
    • FineReader complained about image resolution, which is recorded in EXIF as 72dpi, but it wanted at least 300dpi. This could be handled by fine-tuning during capture, or rewriting EXIF in post.
• FineReader vs. Tesseract
  • FineReader approaches 100% accuracy, though we haven't compared all words yet. Tesseract appears to be in the 70-80% range
  • FineReader groks a page's structure and reproduces it in the resulting PDF. Tesseract doesn't.
  • FineReader includes the background image, which in the previous test is fairly low contrast. It may be distracting for some. Contrast could be bumped up during capture.
  • Tesseract goes straight to text, does not include a background image
  • Tesseract has a library which can be invoked from a scripted scan-OCR solution
  • ABBYY offers to license the library for FineReader to developers, though the price is unknown at this time.
  • FineReader is hands-off, for what it does. After it receives images, it works without prompt or interrupt.

Summary-

• FineReader looks to be a very good product, and is worthy of consideration
• The price for a single station starts at $120
• Getting the same results found in FineReader might take a bunch of developer effort
• Their OCR library might also be gotten at a reasonable cost, given Noisebridge is a non-profit
• Cropping could be handled in a script prior to FineReader ingestion
• As to the question of whether to utilise proprietary softwares, or keep to diy/open-source, can be boiled down to convenience. Time to usable product. With a primary objective of scanning a huge backlog of books, getting high-quality PDFs (or other digital formats) from books very soon is attractive.