Image Analysis of Crystal Size Distribution and Agglomeration for β form L-Glutamic Acid Crystallization based on YOLOv4 Deep Learning

In order to monitor the growth quality of crystal during the crystallization process of β-form L-glutamic acid (LGA), an image analysis method based on YOLOv4 deep learning is proposed for on-line measurement of crystal size distribution (CSD) and detection of crystal agglomeration. Firstly, different shapes of β-form LGA crystal are classified and the corresponding strategies for measuring crystal length are given, respectively. Then, an image segmentation algorithm based on YOLOv4 deep learning is proposed for real-time analysis of in-situ captured images. Meanwhile, YOLOv4 deep learning is adopted to determine the long axis of each segmented crystal image by using a specified threshold and a gradient algorithm, such that the measurement accuracy on crystal length could be effectively improved. By introducing an index named overlapping ratio of bounding box (ORBB), detection of crystal agglomeration is conducted based on classifying the segmented crystal images. Experimental results on β-form LGA crystallization are shown to demonstrate the effectiveness and advantage of the proposed image analysis method.