Mask R-CNN based deep learning analysis on in-situ measured crystal images with automatic dataset labelling

For real-time measurement of crystal size distribution (CSD) by in-situ captured crystal images, a deep-learning based image analysis method is proposed to improve measurement accuracy and efficiency, based on the well recognized mask-regional convolutional neural network (Mask R-CNN). An automatic dataset labelling algorithm is established to facilitate preparing the training dataset that is required to include a large number of crystal image samples for effective deep learning. Moreover, an image thresholding segmentation algorithm is introduced to extract the region of interest (ROI) in each crystal image sample for training the Mask R-CNN, such that improved segmentation accuracy and efficiency could be obtained for online image analysis to measure CSD during a crystallization process. Experimental results on measuring the crystallization process of ß form L-glutamic acid (ß-LGA) are shown to verify the effectiveness and advantage of the proposed method.