Thesis

An investigation of automated visual inspection system design for defect detection

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2025
Thesis identifier
  • T17302
Person Identifier (Local)
  • 201965886
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Non-destructive testing (NDT) methods are in high demand for defect detection across various industries, as they ensure the integrity and safety of materials and products without causing damage. Among NDT techniques, automated visual inspection (AVI) systems have gained significant attention due to their potential for efficient, accurate, and non-invasive defect detection. AVI systems offer a powerful solution by leveraging advanced imaging and artificial intelligence (AI) technologies to identify defects in real time. Based on defect localisation precision, defect detection tasks can be categorised into four levels: image-level, object-level, pixel-level, and 3D-level. Each of these levels necessitates distinct designs and techniques. The key challenges include coarse localisation and limited interpretability at the image level, object segmentation and ambiguity in defining object boundaries at the object level, fine grained localisation and computational complexity at the pixel level, and the integration of spatial information with high computational cost at the 3D level. Despite their potential, several challenges hinder their industrial implementation, including the accuracy-efficiency trade-off, limited computational resources in practical applications, and data scarcity for training deep learning models. This thesis investigates the design and development of AVI systems for defect detection across four different localisation precision levels and diverse industrial applications. Through a systematic design methodology, this research addresses key challenges by optimising system architecture, data acquisition methods, and the detection algorithms to meet the specific requirements of each industrial application. The key components of the AVI system design are comprehensively reviewed, including data acquisition, defect detection, and system design methodologies. Regarding the defect detection methods, a universal perspective on defect detection techniques is provided and the state-of the-art methods are compared. Based on the literature review findings, several knowledge gaps are identified, including the absence of a comprehensive design methodology for AVI systems, challenges in real-time detection on devices with limited computational resources, and under exploration of the anomaly detection methods in AVI systems. The first contribution of the research is to develop a novel comprehensive design methodology to resolve the identified challenges of the AVI system design, with a focus on defect detection method development. Central to this methodology is a design process model, where the design evolves from an abstract and qualitative concept to a specific and quantitative embodiment, and finally a complete AVI system. This design methodology also emphasises the development of defect detection method through a detailed algorithm selection and optimisation process, which can be tailored according to the availability of target data. The second contribution is the implementation of the comprehensive design methodology in developing four AVI systems in different application domains. By adopting the proposed design methodology, four AVI systems are designed and developed in four case studies, targeting different defect localisation precision levels and diverse industrial applications. The first case study emphasises the accuracy-efficiency trade-off in computationally resourceconstrained devices, while the other three case studies address the challenges posed by data scarcity. In the first case study, an image- and pixel-level concrete wall crack detection system is designed, incorporating a novel training strategy for agile development of AI models along with a novel model architecture. The second case study explores object-level detection of wheat head diseases, employing zero-shot learning and domain adaptation techniques to train an anomaly detection model without utilising any disease data. The third case study shifts to pixel-level car engine surface defect detection. It includes a comparative analysis to examine the impact of data collection configurations, anomaly characteristics, and anomaly detection methods. Based on these findings, the most suitable defect detection algorithm is selected and optimised. The fourth case study presents a 3D-level spacecraft anomaly detection system, illustrating the design of both image acquisition system and software platform for a modular spacecraft inspection system. These four case studies collectively demonstrate how the proposed design methodology can be implemented to address distinct requirements and challenges of each industrial application in a systematic manner.
Advisor / supervisor
  • Yan, Xiu-Tian
  • Wong, Andy
Resource Type
DOI

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T17302 2025-06-17 Public Download