OpenBMB has released MiniCPM-OCR 1.0 for reading documents alongside MiniCPM-OCR Jev for evaluating the resulting text. The parser handles Chinese and English documents, including tables and formulas. Jev compares an image region with its recognition result and produces a quality score, including for other parsers’ output. Code, documentation and model weights carry the Apache 2.0 licence.
OpenBMB, the research group behind the MiniCPM model family, released MiniCPM-OCR 1.0 and MiniCPM-OCR Jev on October 27. OCR means optical character recognition, the conversion of text in images into machine-readable text. The release pairs a document parser with a separate model that evaluates recognition results, including output from other parsing systems.[1], [2]
MiniCPM-OCR 1.0 handles Chinese and English documents and PDF files. Its scope includes text, tables, formulas, charts, seals and key information extraction. Scans, photographs, curved pages, skewed images and documents photographed again are among the inputs described by OpenBMB. Parsed content can be returned as Markdown, a text-formatting notation, or JSON, a structured data format that software can process.[1]
Jev receives an image region and its corresponding recognition result, then returns a quality score between 0 and 1 for text, formulas or tables. OpenBMB describes the function as locating content that needs review and distinguishing problematic regions from apparently reliable ones. Its documentation instructs users to set thresholds after checking a sample from the actual application. The project’s code, documentation and both models’ weights are released under Apache 2.0. The parser also has local-deployment and fine-tuning tools.[1]
