Skip to content

Image2Feature: A Framework for Local Machining Feature Recognition in 3D CAD Models via Learning-Based Object Detection from 2D Images

Document Type
Publications · Journal Article
Registered DB
KCISCIEScopus
Journal Ranking
JCR Q1Scopus Q2KCI Q1
Citation
Web of Science 0Scopus 0
FWCI (Scopus)
0
JCR IF
JCR 2024 3.6
ISSN
2234-7593

Authors

  1. Jeon, Ye-SeongFirst authorKumoh Natl Inst Technol, Dept Mech Design Engn, Grad Sch, Gumi, Gyeongbuk, South Korea
  2. Kim, Kun-ChanKumoh Natl Inst Technol, Dept Mech Design Engn, Grad Sch, Gumi, Gyeongbuk, South Korea
  3. Kim, Hyun-ChanCorresponding authorKumoh Natl Inst Technol, Sch Mech Engn, Gumi, Gyeongbuk, South Korea
  4. Kwon, SoonjoCorresponding authorPusan Natl Univ, Sch Mech Engn, Busan, South Korea

Abstract

As digital transformation in the manufacturing industry accelerates, there is growing interest in design and production automation based on three-dimensional (3D) Computer-Aided Design (CAD) models. In this context, Machining Feature Recognition (MFR), which automatically extracts machining information from CAD models, has emerged as a key enabling technology. However, existing Boundary Representation (B-rep)-based methods rely heavily on topological structures, making them prone to failure when recognizing complex or intersecting features. To overcome this limitation, this paper proposes Image2Feature, a framework that combines image-based object detection with B-rep-based geometric validation. The proposed method captures projection and cross-sectional images from 3D CAD models along multiple axes and detects local machining features such as chamfers, fillets, holes, and tapers using a learning-based object detection model. The detection results are mapped to 3D B-rep faces through ray casting, and feature-specific geometric validation rules are applied to produce the final recognition results. In the post-processing stage, composite features such as counterbores, countersinks, and coaxial holes are further identified from the recognized faces. Experiments on 15 test models achieved a mean score of 91.3%. Furthermore, comparative evaluation against AAGNet, BRepNet, and UV-Net was conducted on the MFCAD++ and Fusion 360 Gallery datasets. The proposed method is applicable not only to process planning automation, but also to reverse engineering, design quality inspection, and 3D model retrieval and classification.

Keywords

Machining feature recognitionLocal machining featureComputer-aided designMachine learningObject detectionAUTOMATIC RECOGNITION

Identifiers

Scopus EID
2-s2.0-105049954536
WoS UID
001869656100001