Reproduce “PP-StructureV2: A Stronger Document Analysis System” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chenxia Li, Ruoyu Guo, Jun Zhou, Mengtao An, Yuning Du, Lingfeng Zhu, Yi Liu, Xiaoguang Hu and 1 more
Baidu Inc.'s PP-StructureV2 enhances its predecessor as an intelligent document analysis system, focusing on improving accuracy, efficiency, and practical functionality. The system achieves 11x faster CPU inference for layout analysis, boosts table recognition accuracy by over 4.5 percentage points, and improves key information extraction by up to 9.1% while drastically reducing model sizes and enabling editable Word file output.
Key insights
- Removing the visual feature extraction module from LayoutXLM (creating VI-LayoutXLM) unexpectedly improved Key Information Extraction (KIE) accuracy while drastically reducing model size and inference time.
- Careful architecture optimization using lightweight networks (PP-PicoDet, PP-LCNet) combined with knowledge distillation (FGD, U-DML) significantly boosts CPU inference speed and reduces model size without sacrificing accuracy.
- Addressing text reading order issues (via TB-YX sorting) is critical for improving the performance of downstream tasks like Relation Extraction in KIE.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.