Automatic Labeling of In-Situ Crystal Images based on Lightweight Network Bisenetv2

To enhance the efficiency and accuracy of in-situ crystal image annotation for building up a model to monitor the crystal growth process during cooling crystallization, a deep learning based automatic labeling method is proposed by using the lightweight network Bisenetv2. The image segmentation is conducted by the lightweight network Bisenetv2, for which a hyperparameter optimization (HPO) and class balance training strategy is presented to improve the segmentation accuracy on in-situ captured images. The trained model is converted into the open neural network exchange (ONNX) format to reduce the implemental dependence on any specific or complex environmental configurations. Furthermore, a software package is developed to embed the ONNX format model for implementation. Additionally, a multi-stage image processing algorithm is provided to enhance image quality for analysis, based on the characteristics of in-situ crystal images captured during the crystallization process. Correction algorithms for poorly segmented crystal contours are also designed, resulting in a final mean intersection over union (mIOU) of 0.8216, with the automated annotation process taking approximately 1 second per image. Experimental results on in-situ images of $\beta$ form L-glutamic acid ($\beta$-LGA) captured from different batches of crystallization validate the effectiveness and robustness of the proposed method.