🤖 AI Summary
Baidu has launched Unlimited-OCR-2, a small document vision-language model available as open weights on ModelScope, designed to read and interpret documents with intricate formatting, such as PDFs from medical lab reports. The model employs Sparse Reference Sliding Window Attention (SR-SWA) to efficiently manage long document decoding without incurring the high computational costs associated with traditional full attention methods. In practical tests, the model demonstrated its capabilities by successfully extracting structured data from complex hospital documents, including multi-language headers and tables.
However, it struggles with dual-panel printouts where result lists are presented side-by-side within a single rectangle. This layout presents a challenge as Unlimited-OCR-2 attempts to merge two separate result lists into one table, often leading to data misalignment. To address this, users can implement two fixes: splitting the document into left and right crops for independent OCR processing and applying a custom parser to reorganize the misaligned output. While the model performs well in extracting labeled data, its limitations underline the need for additional coding and customized solutions to fully utilize its capabilities in varied formatting scenarios, especially in medical environments.
Loading comments...
login to comment
loading comments...
no comments yet