Optimized Methods for Online Monitoring of L-Glutamic Acid Crystallization

Abstract

In order to monitor the crystallization process of L-glutamic acid online, a real-time detection method based on non-invasive image analysis has been proposed to obtain in-situ images, and a deep-learning based network Mask R-CNN is applied to detect target crystals in images. Considering deep-learning network requires an enormous amount of dataset with labelled region of interest (RoI) samples, this paper proposes semi-automatic labelling methods to reduce human work when generating the dataset. By applying another Mask R-CNN for labelling the dataset, human work can be reduced from labelling the whole dataset to filtering the detection results of the labeller Mask R-CNN. The final detection results prove the feasibility of this method. The proposed method is also proved to be more feasible and reliable than transfer learning.