Deep learning for imaging particle analysis applications

This work offers a thorough exploration of an innovative methodology for imaging particle analysis, utilizing advanced deep learning (DL) techniques. Chapter 1 points out the significant role of powders and particles in various industries, with a focus on the importance of particle size distributions (PSDs) for product and process quality. It explains the advantage of imaging particle analysis over other methods for detailed size and shape measurement, highlighting the limitations of manual and traditional automated techniques and ultimately advocates for DL as a promising approach to enhance imaging particle analysis, aiming for a more reliable, objective, and precise analysis process that reduces manual labor and error susceptibility. Chapter 2 reviews the state of the art of imaging particle analysis techniques, ranging from manual counting methods and traditional automated techniques like Otsu’s method, k-means, watershed transform and Hough transform, to classical machine learning (ML)-based methods that laid the foundation for the work at hand. Chapter 3 establishes a basis for understanding particle properties and concepts relevant to the studies conducted in the course of this work, such as the concept of equivalent diameters, the PSD and sintering. Chapter 4 delves into the theoretical underpinnings of ML, DL, and computer vision, offering a primer on the technologies that form the core of this work. It discusses neural network architectures and their applications, focusing on the potential of convolutional neural networks (CNNs) and region-based convolutional neural networks (R-CNNs), such as the Feature Pyramid Network and Mask R-CNN, for imaging particle analysis. Chapter 5 focuses on the adaptation of the aforementioned DL methodologies, especially R-CNNs, for particle analysis. It showcases the customization of DL techniques to address specific challenges in particle analysis, such as enhancing training performance and improving PSD prediction accuracy. Furthermore, two novel extensions of already existing DL methods – Voronoi++ for the analysis of densely packed grains in sintered solids and FibeR-CNN for the analysis of curved and overlapping fibers – are introduced. Chapter 6 presents the synthPIC (Synthetic Particle Image Creator) toolbox, a novel solution for overcoming the data-intensive requirements of training DL models by generating synthetic training images. This approach boosts the potential for DL in imaging particle analysis by sidestepping the limitations associated with gathering vast amounts of manually annotated data. Chapter 7 presents real-world and synthetic datasets used for the training and evaluation of the proposed DL particle analysis methods. It emphasizes the diversity of the datasets and the challenges they present, underscoring the necessity and effectiveness of the developed DL approaches in overcoming these challenges. Chapter 8 provides a detailed evaluation of the proposed DL methodologies through five case studies: detection of complete soot aggregates in transmission electron microscopy images; detection of quasi-spherical primary particles in secondary electron microscopy (SEM) images of silica aggregates; classification of catalyst particles in endoscopic images of a densely packed binary particle mixtures; detection of packed crystallites in SEM images of breaking edges of solids made from compacted and sintered zinc oxide particles; and detection of curved and overlapping fibers in SEM images of carbon nanotubes. By comparing the utilized DL techniques to traditional automated and manual imaging particle analysis methods, an improved accuracy and reliability across the various evaluated applications is demonstrated. Finally, Chapter 9 summarizes the contributions and implications of this work, while acknowledging its limitations and potential societal impact, as well as pointing out possible directions of future work.