In-line image analysis of particulate processes with deep learning: Optimizing training data generation via copy-paste augmentation

Abstract

Monitoring particle properties directly in the process using in-situ microscopes can provide valuable input for control loops, improve process understanding, and facilitate process optimization. However, obtaining reliable information from these images is still challenging, particularly for higher solids concentrations or agglomerating systems. Recent studies have successfully applied deep learning models to extract particle characteristics from in-situ image data. Despite these advances, the problem of generating training data has not been properly addressed. Manual annotation is time-consuming and prone to bias due to high particle counts, particle overlap, and out-of-focus objects. This paper introduces a new approach to generating training data for segmentation models by combining conventional segmentation methods with copy-paste augmentation. A case study was conducted in which depth filtration experiments were performed using irregularly shaped alumina particles. An instance segmentation model trained on data created with the proposed approach successfully detected and characterized particles, even at high solids concentrations.