kth.sePublications KTH
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Towards Automated Assembly Quality Inspection with Synthetic Data and Domain Randomization
KTH, School of Electrical Engineering and Computer Science (EECS), Intelligent systems, Robotics, Perception and Learning, RPL. Scania CV AB, Sweden.ORCID iD: 0000-0002-4180-3809
Scania CV AB, Sweden.
Scania CV AB, Sweden.
Scania CV AB, Sweden.
Show others and affiliations
2025 (English)In: Proceedings: IEEE/CVF International Conference on Computer Vision Workshop, ICCVW, 2025, 2025, p. 1395-1403Conference paper, Published paper (Refereed)
Abstract [en]

Assembly quality inspection plays a vital role in manufacturing, where correct part placement and alignment directly affect product reliability. While deep learning–based object detection offers a promising solution for automatic assembly quality inspection, it is hindered by data scarcity. Training on synthetic data with Domain Randomization (DR) helps address this challenge, yet existing DR methods focus on generating individual objects and do not capture the relational structure needed for assembly inspection. In this paper, we identify two key factors for effective synthetic data generation in assembly inspection: preserving spatial relationships between components and providing part-level textures and annotations. We propose an Assembly-Specific Generation Scheme that incorporates these factors into a state-of-the-art DR pipeline. To evaluate its impact, we introduce SIP2A-OD, a new object detection dataset comprising two real-world assembly use cases collected under varied manufacturing conditions. We train a YOLOv12 model on synthetic data generated by our pipeline and test it on real data from the SIP2A-OD dataset. Compared to the baseline pipeline designed for individual object detection, our method improves mAP@50 by more than 15% in both use cases. These results demonstrate the effectiveness of our scheme and its potential for broader applications in industrial assembly inspection without the need for manual data collection or annotation.

Place, publisher, year, edition, pages
2025. p. 1395-1403
National Category
Engineering and Technology
Identifiers
URN: urn:nbn:se:kth:diva-372634OAI: oai:DiVA.org:kth-372634DiVA, id: diva2:2012933
Conference
IEEE/CVF International Conference on Computer Vision Workshop 2025, Honolulu, Hawaii, USA, October 19-25, 2025
Note

QC 20251112

Available from: 2025-11-11 Created: 2025-11-11 Last updated: 2025-11-12Bibliographically approved
In thesis
1. Towards Automated Parts Recognition in Manufacturing with Synthetic Data
Open this publication in new window or tab >>Towards Automated Parts Recognition in Manufacturing with Synthetic Data
2025 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

This thesis advances the understanding and application of synthetic data for manufacturing parts recognition. Vision-based inspection systems in manufacturing rely heavily on real image data, which are costly to collect, annotate, and adapt across products and environments. To address these challenges, this work presents a systematic investigation of how synthetic data can be effectively generated, evaluated, and applied for robust and scalable performance. The research introduces a series of new industrial benchmark datasets covering multiple manufacturing use cases and factory environments: SIP-17, SIP15-OD, and SIP2A-OD, to enable unified evaluation of sim-to-real transfer in classification and detection tasks. Building on these datasets, a domain randomization pipeline is developed to systematically explore the effects of rendering parameters, material variability, and illumination on model generalization. To further automate data generation, the thesis proposes Synthetic Active Learning (SAL), a closed-loop framework that identifies model weaknesses and adaptively refines synthetic data generation without requiring real samples or manual tuning. Experiments across the benchmark datasets show that the proposed method improves model robustness compared to existing approaches while reducing manual labeling requirements. Collectively, these contributions provide new insights into how synthetic data can be systematically leveraged to build data-efficient, automated, and reliable vision systems for manufacturing, aiming to support future development of intelligent and flexible production systems.

Abstract [sv]

 Denna avhandling bidrar till förståelsen och tillämpningen av syntetisk data för igenkänning av tillverkningskomponenter. Visionsbaserade inspektionssystem inom industrin är starkt beroende av verkliga bilddata, vilka är kostsamma att samla in, annotera och anpassa mellan olika produkter och miljöer. För att hantera dessa utmaningar presenterar arbetet en systematisk undersökning av hur syntetisk data effektivt kan genereras, utvärderas och tillämpas för robust och skalbar prestanda. Forskningen introducerar en serie nya industriella benchmark-datamängder som täcker flera användningsfall och fabriksmiljöer: SIP-17, SIP15-OD ochSIP2A-OD, för att möjliggöra en enhetlig utvärdering av sim-to-real-överföring inom klassificerings- och detektionsuppgifter. Med utgångspunkt i dessa datamängder utvecklas en domain randomizationpipeline som systematiskt undersöker effekterna av renderingsparametrar, materialvariation och belysning på modellens generaliseringsförmåga. För att ytterligare automatisera datagenereringen föreslås Synthetic Active Learning (SAL), ett slutet ramverk som identifierar modellens svagheter och adaptivt förfinar den syntetiska datagenereringen utan att kräva verkliga prover eller manuell justering. Experiment över benchmark-datamängderna visar att de föreslagna metoderna förbättrar modellens robusthet jämfört med befintliga tillvägagångssätt, samtidigt som behovet av manuell annotering minskar. Sammantaget ger dessa bidrag nya insikter i hur syntetisk data systematiskt kan utnyttjas för att bygga dataeffektiva, automatiserade och tillförlitliga visionssystem för tillverkningsindustrin, med målet att stödja utvecklingen av framtida intelligenta och flexibla produktionssystem.

Place, publisher, year, edition, pages
Stockholm: KTH Royal Institute of Technology, 2025. p. xix, 75
Series
TRITA-EECS-AVL ; 2025:105
Keywords
Synthetic Data, Sim-to-real Transfer, Object Detection, Domain Randomization, Ac tive Learning, Dataset Generation, Automation, Computer Vision, Industrial Inspec tion, Syntetisk Data, Sim-to-real-överföring, Objektigenkänning, Domain Randomization, Aktivt Lärande, Datagenerering, Automation, Datorseende, Industriell Inspektion
National Category
Computer Sciences
Identifiers
urn:nbn:se:kth:diva-372644 (URN)978-91-8106-466-7 (ISBN)
Public defence
2025-12-09, https://kth-se.zoom.us/j/62913476523, Kollegiesalen, Brinellvägen 8, Stockholm, 13:00 (English)
Opponent
Supervisors
Funder
Wallenberg AI, Autonomous Systems and Software Program (WASP)
Note

QC 20251112

Available from: 2025-11-12 Created: 2025-11-11 Last updated: 2025-11-24Bibliographically approved

Open Access in DiVA

No full text in DiVA

Other links

fulltext

Authority records

Zhu, XiaomengBjörkman, MårtenMaki, Atsuto

Search in DiVA

By author/editor
Zhu, XiaomengBjörkman, MårtenMaki, Atsuto
By organisation
Robotics, Perception and Learning, RPL
Engineering and Technology

Search outside of DiVA

GoogleGoogle Scholar

urn-nbn

Altmetric score

urn-nbn
Total: 70 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf