Optical Character Recognition

The development of computer vision and deep learning is pushing OCR solutions into the AI limelight. The current technological possibilities provide a broad spectrum of business applications, including digitization of handwritten documents, analysis of text in photos, and sign recognition by autonomous vehicles. At deepsense.ai we use both open-source solutions and our own models in OCR-oriented projects.

Open-source or commercial solutions are not always able to meet OCR-specific business needs. Harnessing the full potential of this technology requires a combination of both open-source solutions and dedicated models.

The wide range of OCR use cases includes the reading of:

• documents (machine font, grammar)
• code (alphanumeric strings)
• labels in natural surroundings (street signs, for example)
• logos
• handwritten text
• and text detection (indicating where text is)

All this reading can be done not only on different surfaces (paper, metal, glass) but also across different printing techniques (printing, engraving, punching), light conditions and fonts.

At deepsense.ai, we have expanded our OCR capabilities with additional pre- and post-processing functionalities. These complement the new, dedicated models we train to accurately address unconventional cases. Such cases include identifying small pieces of text, vertical text, different font sizes and text appearance on various materials. For a recent R&D project we developed a dedicated, fully-convolutional neural net architecture that quickly scans dense and complex printed diagrams.

The system detects and reads hundreds of small codes and labels with 93% accuracy.