A method and related device for manufacturing feature analysis and interaction of a two-dimensional engineering drawing

By constructing a million-level bimodal annotation dataset and a multimodal fusion model, the problem of visual and vector separation in two-dimensional engineering drawings was solved, achieving high-precision manufacturing feature analysis and intelligent interaction, and improving the accuracy and consistency of recognition and positioning.

CN122369052APending Publication Date: 2026-07-10ZHEJIANG QINGBEI INTELLIGENT TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
ZHEJIANG QINGBEI INTELLIGENT TECHNOLOGY CO LTD
Filing Date
2026-04-07
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing technologies cannot effectively analyze manufacturing features in two-dimensional engineering drawings. Pure visual methods lack geometric accuracy, and pure vector methods cannot understand semantic content. The separation between visual and vector methods makes integration difficult.

Method used

We construct a dataset of millions of industrial 2D engineering drawings for bimodal annotation, use graph convolutional networks for vector topological semantic pre-classification, combine a rotating target detection model and a multimodal fusion visual language model to achieve deep fusion of vision and vector, refine the boundaries through bidirectional mapping and mutual enhancement mechanisms, and output structured manufacturing feature semantic data.

Benefits of technology

It significantly improves the accuracy and positioning precision of manufacturing feature analysis, more than doubles the accuracy of identifying extremely sparse regions, improves positioning precision to the vector level, achieves an F1 score of 98.6%, and enables intelligent interactive editing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122369052A_ABST
    Figure CN122369052A_ABST
Patent Text Reader

Abstract

This invention discloses a method and related apparatus for manufacturing feature analysis and interaction of two-dimensional engineering drawings, belonging to the field of computer vision technology. The method includes: constructing a million-level bimodal labeled dataset; simultaneously acquiring vector graphic data and high-resolution raster images and establishing spatial coordinate mapping; using graph convolutional networks for vector topological semantic pre-classification; using a rotated target detection model to locate manufacturing feature regions; establishing a closed-loop mutual enhancement mechanism between vision and vector to dynamically adjust confidence and correct boundaries; and outputting structured manufacturing feature semantic data through a multimodal fusion visual language model. This invention overcomes the recognition difficulties caused by pixel sparsity in pure vision methods through deep fusion of visual and vector features, improving the recognition accuracy by more than double in extremely sparse regions, and achieving an F1 score of 98.6% for 15 categories of manufacturing feature analysis.
Need to check novelty before this filing date? Find Prior Art