Context-Aware Surface Defect Detection in Reflective Objects

 This Ph.D. thesis presents novel approaches to surface defect inspection of narrow and reflective objects, addressing the limitations of traditional manual methods and enhancing vision-based detection techniques. While manual inspections are valuable, they are susceptible to human errors and inefficiencies. Moreover, despite their potential, automated vision-based methods face significant challenges due to the need for high-quality datasets and adaptability in complex lighting and background conditions.

Performance issues and real-world environmental complexities pose significant barriers to the practical deployment of these models. Traditional quality assurance mechanisms often struggle to accommodate the intricacies of evolving deep learning models. This research aims to bridge this gap by comprehensively understanding and validating these tools for real-world integration.

Our methodology unfolds in four main stages. Initially, we employ data-based strategies to improve defect detection performance. The first major step involves the implementation of our model-based Image Quality Assessment (IQA) algorithm to refine defect detection capabilities, improving the assessment and categorization of surface damage. Following this, we add a key feature by incorporating precise segmentation based on the object's Building Information Modeling (BIM), which enhances image evaluation with contextual and structural insights for more accurate defect analysis. Finally, we developed an IQA-based viewpoint system to optimize data acquisition in complex environments, strategically capturing high-quality images even in challenging conditions.

Central to our strategy is a deep learning model proficient in defect classification, localization, and segmentation. Image Processing Techniques (IPTs) enable it to handle complex subjects like window frames using legged robotic platforms. Empirical testing demonstrates the system's ability to identify subtle defects, particularly in environments with variable lighting. This work is envisioned to scale to a Machine Vision system, broadening industrial applicability and enabling comprehensive asset monitoring and evaluation.

Our methods underscore the significant impact of IPTs, IQA, and segmentation on enhancing defect detection accuracy, especially for reflective and narrow architectural objects such as window frames. These innovative vision-driven strategies yielded remarkable improvements in defect precision, highlighting the value of data-driven enhancement techniques in challenging environments.

Overall, this thesis offers an integrated and efficient solution to surface damage detection in industrial settings, focusing on accuracy, efficiency, and practical application. It seeks to modernize quality inspection practices, significantly advancing industrial maintenance standards and safety. Future integration of these four methods into a machine vision system on the Spot robot holds promise for further enhancing defect detection and inspection capabilities.