U-net based Deep-Learning Image monitoring of Crystal Size Distribution during L-Glutamic Acid Crystallization

Abstract

For monitoring the crystal size distribution (CSD) during L-glutamic acid (LGA) crystallization process, a deep-learning image analysis method is proposed based on the U-net network. The proposed method consists of crystal image preprocessing, crystal image segmentation, and CSD measurement. To cope with the negative influence on image analysis from continuous stirring of solution and particle motion in a crystallizer, the U-net network is adopted to perform deep learning for crystal image segmentation, with no need to prepare a large amount of training samples of real crystal images. Moreover, this approach could be implemented in real time for analysis of in-situ captured microscopic images, thus facilitating on-line measurement of CSD. Consequently, a log-normal distribution model is established for effectively depicting the one-dimensional (i.e. length) distribution of β-form LGA crystals during crystallization. A numerical example and experimental study on the cooling crystallization process of β-form LGA are shown to demonstrate the effectiveness the proposed method.