PCB Design Document Extraction and Table Completion Method with YOLOv11 and Multi-Agent Architecture
Zongbin Qiao, Xiaozhi Du, Jinhua Yang, Hairui Liu · 2025
In the PCB design process, the Datasheet is an essential reference document for designers, containing key parameters such as pin configurations. However, current OCR and multimodal models still face challenges in document extraction, especially the loss and inaccurate recognition of table data, which is particularly evident when table lines are incomplete. To address these issues, this paper proposes an innovative method based on the YOLOv11 model, which can precisely extract the missing parts of a table. First, the YOLOv11 model is trained to automatically identify and extract incomplete lines from tables. Then, an algorithm developed by the authors is used to complete these missing table lines. The algorithm determines the reference line by statistical analysis of vertical pixels, binarization processing, and convolution operations, simplifying the image according to established rules, thus achieving clear division and segmentation between columns. To enhance the efficiency of document extraction, this paper also designs a multi-agent multimodal large model architecture. Specifically, one agent extracts the full text, another summarizes the table data and outputs it in Markdown format, and the final agent cross-checks and combines the table information with the full text, ultimately outputting a plain text result. This system will be integrated into a graph database in the future to provide high-quality document data support for large models, promoting further development of intelligent applications.