Automatic detection, segmentation, and characterization of 3-nitro-1,2,4-triazol-5-one (NTO) crystal particles during cooling crystallization using deep learning

The complex cooling crystallization process of 3-nitro-1,2,4-triazol-5-one (NTO) directly affects its performance. The introduction of process analysis technology is expected to provide an effective solution for process monitoring and optimization. This study proposes a crystal instance segmentation method based on the Mask R-CNN framework and develops a semi-automatic annotation strategy combining model-assisted labeling and manual refinement. Additionally, image augmentation techniques were used to enhance model stability. Comparative experiments show that the model trained with data augmentation and semi-automatic annotation achieved an AP50 of 94.4% on the validation set, improving annotation efficiency by approximately 12-fold. The model was successfully applied to the automatic segmentation and parameter extraction of NTO crystallization images, enabling effective quantitative analysis of particle count, size distribution, and morphology. This approach offers an efficient technical solution for online image monitoring of complex crystallization systems, with significant potential for industrial application. • Mask R-CNN was used for instance segmentation of NTO crystallization images. • A semi-automatic labeling strategy improved annotation efficiency and accuracy. • Data augmentation boosted model performance in multi-scale particle detection. • Crystal morphology was quantitatively analyzed during NTO crystallization.