Deep Learning based OCR System
Printed text is everywhere, from financial documents to serial identification codes, and the often first step to digital transformation of your business is digitizing this printed text. SaigeOCR® employs the latest advances in deep learning to automatically detect the location of words and text in images and to transform it into digital strings. SaigeOCR® extracts curvature and other geometric information of the region surrounding the text, and robustly recognizes and transforms the txt into digital strings taking into account any distortions. SaigeOCR® offers state-of-the-art performance that delivers an efficient, real-time, robust OCR solution.
1. Fast, lightweight, and robust detection using geometric information
- Unlike existing commercial OCR products that focus on detecting the location of the text, SaigeOCR® extracts geometric information of the text region such as rotational and translational displacements, distortion, and curvature. SaigeOCR® can then “undistort” the surrounding region to straighten out any curved or distorted text, allowing it to robustly recognize arbitrarily distorted text.
- SaigeOCR®’s lightweight network architecture delivers real-time OCR, making it ideal for high-volume applications like manufacturing and document processing.
2. Powerful OCR database
Virtual OCR dataset generation technology
- Different industries have different text formats and printing conventions. For example, financial documents adhere to a standard format, while in logistics unconstructed text can be printed on the surfaces of products.
- SaigeOCR® uses our proprietary OCR text database to pretrain the network with a range of clean and distorted text samples. The network is then fine-tuned to detect varieties of text in different target domains.
3. Document understanding
- Using both text string and location information obtained from OCR, SaigeOCR® automatically extracts and reformats the text string information into a user-specified structure.
- Combining the OCR and document understanding capabilities, SaigeOCR® outputs, from any unstructured image, the desired information into a user-specified structure.