Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
On-edge assembly defect detection in noisy environments using convolutional neural networks
Smit, J. 2025. On-Edge Assembly Defect Detection in Noisy Environments using Convolutional Neural Networks. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/07ed5803-7c34-4d1f-add1-9430dd0a134c
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613740270092288 |
|---|---|
| access_status_str | Open Access |
| author | Smit, Janre |
| author2 | Burger, Leon E. |
| author_browse | Burger, Leon E. Smit, Janre |
| author_facet | Burger, Leon E. Smit, Janre |
| author_sort | Smit, Janre |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Smit, J. 2025. On-Edge Assembly Defect Detection in Noisy Environments using Convolutional Neural Networks. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/07ed5803-7c34-4d1f-add1-9430dd0a134c |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/132428 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:40:56.936Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2025 |
| publishDateRange | 2025 |
| publishDateSort | 2025 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/132428 On-edge assembly defect detection in noisy environments using convolutional neural networks Smit, Janre Burger, Leon E. Schutte, Corne S. L. Stellenbosch University. Faculty of Engineering. Dept. of Industrial Engineering. Neural networks (Computer science) Manufacturing processes -- Automation Image processing -- Digital techniques Augmented reality -- Industrial applications UCTD Smit, J. 2025. On-Edge Assembly Defect Detection in Noisy Environments using Convolutional Neural Networks. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/07ed5803-7c34-4d1f-add1-9430dd0a134c Thesis (MEng)--Stellenbosch University, 2025. ENGLISH ABSTRACT: This thesis investigates the use of convolutional neural networks (CNNs) for defect detection in assembly environments, focusing on deployment in noisy settings and operation on edge devices, specifically the NVIDIA Jetson Nano module. The research addresses a gap in existing defect detection methods, aiming to ensure robust performance under visually noisy conditions. Key research questions revolve around the accuracy of CNNs in noisy environments, the effectiveness of image preprocessing and data augmentation, and the feasibility of deploying CNN-based systems on resource-constrained edge devices. A major contribution is the development of a specialised dataset of real-world assembly images to assess CNN robustness under challenging conditions. Various data augmentation techniques, including pixel-level annotation, were explored to improve model performance, particularly in noisy environments. The results showed that CNN-based object detection, especially with fine-grained pixel-level annotations, outperformed traditional image classification models in detecting defects in noisy assembly images. The thesis also demonstrates the successful deployment of a CNN-based defect detection model on the NVIDIA Jetson Nano, evaluating the feasibility of real-time, edge-based defect detection in industrial environments. This confirms that edge devices like the Jetson Nano are viable platforms for CNN-based defect detection, even in resource-constrained settings, enabling real-time quality assurance solutions without relying on centralised computing resources. A peer-reviewed conference article titled ”CNN Defect Detection Using Image Classification Falls Short in Assembly” was developed from this thesis and presented at the 2024 International Conference on Intelligent and Innovative Computing Applications (ICONIC 2024). The article examines the limitations of image classification CNNs for defect detection in assembly processes, evaluating several pre-trained CNNs and a custom image classification CNN on a dataset of model train seat assembly images. Despite data augmentation and image processing techniques, the models struggled to accurately predict defects, especially in noisy environments. Grad-CAM analysis revealed that the models focused on irrelevant features, highlighting the limitations of image classification CNNs for defect detection in noisy assembly environments. The findings contribute to the field of industrial defect detection, offering insights into the challenges and solutions for implementing machine learning models in noisy, real-world environments. The work advances the understanding of optimising CNNs for assembly settings and provides a practical approach for deploying these models on edge devices for real-time applications. Future research directions include expanding defect detection capabilities, integrating more complex CNN architectures, and exploring additional edge devices for enhanced scalability and performance. AFRIKAANSE OPSOMMING: Hierdie tesis ondersoek die gebruik van ’convolutional neural networks’ (CNNs) vir die opsporing van defekte in samestellingsomgewings, met ’n fokus op implementering in omgewings met ho¨e vlakke van visuele geraas en werking op ’edge’-toestelle, spesifiek die NVIDIA Jetson Nano-module. Die navorsing spreek ’n gaping in bestaande defekdeteksie-metodes aan en poog om robuuste werkverrigting in visueel raserige omstandighede te verseker. Belangrike navorsingsvrae fokus op die akkuraatheid van CNNs in raserige omgewings, die doeltreffendheid van beeldvoorverwerking en data-augmentering, asook die haalbaarheid van die implementering van CNN-gebaseerde stelsels op hulpbron-beperkte ’edge’-toestelle. ’n Groot bydrae van die werk is die ontwikkeling van ’n gespesialiseerde databasis van regte-wˆereld foto’s van ’n produksielynsamestelling om die robuustheid van CNNs onder uitdagende toestande te evalueer. Verskeie data-augmenteringtegnieke, insluitend pixelvlak-aanmerkings, is ondersoek om modelle se werkverrigting, veral in raserige omgewings, te verbeter. Die resultate toon dat CNN-gebaseerde objekdeteksie, veral met ’fine-grained’ pixelvlak-aanmerkings, beter presteer het as tradisionele ’image classification’-modelle in die opsporing van defekte in raserige samestellingsbeelde. Die tesis toon ook die suksesvolle implementering van ’n CNN-gebaseerde defekdeteksie-model op die NVIDIA Jetson Nano, wat die haalbaarheid van regstreekse, ’edge’-gebaseerde defekdeteksie in industri¨ele omgewings evalueer. Dit bevestig dat ’edge’-toestelle soos die Jetson Nano lewensvatbare platforms vir CNN-gebaseerde defekdeteksie is, selfs in hulpbron-beperkte omgewings, wat regstreekse kwaliteitsversekering-oplossings moontlik maak sonder om op gesentraliseerde ekenaarbronne staat te maak. ’n ’Peer-reviewed’ konferensie-artikel getiteld ”NN Defect Detection Using Image Classification Falls Short in Assembly” is vanuit hierdie tesis ontwikkel en by die 2024 International Conference on Intelligent and Innovative Computing Applications (ICONIC 2024) aangebied. Die artikel ondersoek die beperkings van ’image classification’ CNNs vir defekdeteksie in samestellingsprosesse, en evalueer verskeie vooropgeleide CNNs sowel as ’n op unieke ’image classification’ CNN op ’n databasis van foto’s van modeltrein sitplek samestellings. Ten spyte van data-augumentering en beeldverwerkingstegnieke, het die modelle gesukkel om defekte akkuraat te voorspel, veral in raserige omgewings. Grad-CAManalise het getoon dat die modelle op irrelevante kenmerke gefokus het, wat die beperkings van ’image classification’ CNN vir defekdeteksie in raserige samestellingsomgewings beklemtoon. Die resultate dra by tot die veld van industri¨ele defekdeteksie en bied insigte in die uitdagings en oplossings vir die implementering van masjienleer-modelle in raserige, regte-wˆereld omgewings. Die werk bevorder die begrip van die optimalisering van CNNs vir samestellingsomgewings en verskaf ’n praktiese benadering vir die implementering van hierdie modelle op ’edge’-toestelle vir regstreekse toepassings. Toekomstige navorsingsrigtings sluit die uitbreiding van defekdeteksie-verwante vermo¨ens in, die integrasie van meer komplekse CNN-argitekture, en die verkenning van addisionele ’edge’- toestelle vir verbeterde skaalbaarheid en werkverrigting. Masters 2025-06-06T11:35:43Z 2025-06-06T11:35:43Z 2025-03 Thesis https://scholar.sun.ac.za/handle/10019.1/132428 en Stellenbosch University 150 pages : ill. application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Neural networks (Computer science) Manufacturing processes -- Automation Image processing -- Digital techniques Augmented reality -- Industrial applications UCTD Smit, Janre On-edge assembly defect detection in noisy environments using convolutional neural networks |
| title | On-edge assembly defect detection in noisy environments using convolutional neural networks |
| title_full | On-edge assembly defect detection in noisy environments using convolutional neural networks |
| title_fullStr | On-edge assembly defect detection in noisy environments using convolutional neural networks |
| title_full_unstemmed | On-edge assembly defect detection in noisy environments using convolutional neural networks |
| title_short | On-edge assembly defect detection in noisy environments using convolutional neural networks |
| title_sort | on edge assembly defect detection in noisy environments using convolutional neural networks |
| topic | Neural networks (Computer science) Manufacturing processes -- Automation Image processing -- Digital techniques Augmented reality -- Industrial applications UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/132428 |
| work_keys_str_mv | AT smitjanre onedgeassemblydefectdetectioninnoisyenvironmentsusingconvolutionalneuralnetworks |