Loading...
Computer vision deep learning methodology for defect detection on resource constrained devices
Date
2025
Abstract
The fourth industrial revolution aims to enhance data-driven decision-making and introduce the adaptability to upgrade the deployed systems in manufacturing dynamically. This revolution also merges physical and digital technologies through the use of data systems and artificial intelligence. Over the decade, state-of-the-art computer vision deep learning algorithms have become popular in detecting defects in manufactured products. However, an industry-focused guideline to build an automated, edge-based defect detection system from a limited target dataset is often not considered.
This thesis describes a body of research on a data-driven, factory-focused Computer Vision-Deep Learning (CV-DL) methodology approach for robust defect detection in manufacturing. The baseline analysis of the state-of-the-art computer vision models on the embedded edge proxy device against the GPU-accelerated device (concerning accuracy and inference time) was conducted with person and face-mask detection experiments. The work then uses the CV-DL methodology in mobile phone defect detection for a remanufacturing use case. This work contributed a mobile phone defect detection dataset and proposed criteria to determine a novel overall grade for the returned mobile phone based on the defects. The baseline results for various YOLO models trained on this dataset are presented in this thesis. This work also presents a pre-fine-tuning experiment (using related datasets) that improved the existing detection accuracy for severe crack-line defects.
An ultrasonic welding (USW) experiment using the CV-DL methodology to develop a weld defect detection and strength prediction sys tem is proposed, overcoming the challenge of an extremely limited dataset using offline and online data augmentations. Convolutional autoencoders, used in this work, extracted the image data, which was combined with input parameter data to build a multimodal dataset for weld strength prediction. The results from this experiment lead to the broader prospect of combining multiple modalities of data for process monitoring and improvement of performance in manufacturing. The CV-DL weld defect detection system was trained on the USW defect dataset developed and annotated in this work. The results from this experiment were benchmarked against a publicly available surface detection dataset.
Finally, a comprehensive study of existing incremental learning techniques was investigated as a potential future research direction to upgrade an existing model with a new class. Incremental learning was successfully implemented on the mobile phone defect detection dataset with minimal evidence of catastrophic forgetting. The replay method was used during retraining to alleviate catastrophic forgetting. The results present an updated model with a new class added, which indicates the significance of incremental learning as a future research direction for defect detection.
Supervisor
Hayes, Martin J.
Southern, Mark
Southern, Mark
Description
Publisher
University of Limerick
Citation
Collections
Files
ULRR Identifiers
Funding code
Funding Information
Sustainable Development Goals
External Link
Type
Thesis
Rights
http://creativecommons.org/licenses/by-nc-sa/4.0/