Deep-learning based in-situ micrograph analysis of high-density crystallization slurry using image and data enhancement strategy

Abstract

Process analytical technology (PAT) plays a crucial role in optimizing crystalline powder product qualities, improving process stability, and reducing experimental costs by enabling real-time monitoring and control of process variables during the crystallization process. In-situ process imaging and analysis have gained significant attention due to their capacity to provide abundant information through images, leading to remarkable advancements in image processing technology, particularly with the support of artificial intelligence. However, the performance suffers when processing high-density slurry images due to challenges such as defocusing, overlapping, background and crystal edges blurriness, and inconsistencies in crystal scales. To address these challenges and explore research strategies based on deep learning for segmenting and analyzing high slurry density images effectively, this study constructed a specific dataset that contains high-density slurry crystal images which is well-labeled, and introduced state-of-the-art neural networks including YOLOv8, U2-net, and Mask R-CNN combined with image and data enhancement strategies. Comparative analysis revealed that YOLOv8 outperformed Mask R-CNN and U2-net by capturing multi-dimensional information in high-density crystal slurry scenarios. Finally, based on the segmentation results of the practical taurine crystallization process using proposed strategy, 1D crystal size, aspect ratio along with 2D size distributions were extracted for further evaluation of crystallization kinetics and crystal properties accurately and quickly.