We proposed a deep learning-based in situ microscopic image analysis system for detecting particles and performing size analysis in a high-density slurry, which shows great potential usage in the area of solution crystallization process. A cost-effective imaging system consisting of a flow-through cell and a 3D-printed microscopic probe was built for high-quality image acquisition. The state-of-the-art deep learning model, Mask RCNN, was used to segment the overlapping particles and classify their categories with high accuracy. A comprehensive performance evaluation of the proposed system was conducted including extrapolation to unseen particle scale, detection in different solids concentration levels, and separation of two different types of particles. Compared with the previous studies, the solids concentration detection limit was improved by five times higher in terms of particle number per frame and three times higher regarding the particle pixel fill ratio (PFR). The categorized detections successfully classified the two different particles in a mixed suspension, and the individual particle size information was extracted, which showed high consistency with the particle information. What's more, a progressive labelling strategy was employed to improve the processing efficiency and accuracy, which would enable the transfer application in solution crystallization process for various crystal species.

我们提出了基于深度学习的原位显微图像分析系统，用于在高密度浆液中检测颗粒并进行尺寸分析，这在溶液结晶过程领域显示出巨大的潜力。构建了一个由流通池和3D打印的显微探针组成的经济高效的成像系统，以获取高质量的图像。最先进的深度学习模型Mask RCNN用于分割重叠的粒子并对其分类进行高精度分类。对提出的系统进行了全面的性能评估，包括外推到看不见的颗粒规模，检测不同固体浓度水平以及分离两种不同类型的颗粒。与以前的研究相比，就每帧颗粒数而言，固体浓度检测极限提高了五倍，而颗粒像素填充率（PFR）则提高了三倍。分类检测成功地将混合悬浮液中的两种不同颗粒进行了分类，并提取了各自的粒径信息，这与颗粒信息具有高度一致性。此外，采用了渐进标记策略来提高处理效率和准确性，这将使转移应用能够应用于各种晶体的溶液结晶过程中。提取出各个粒径信息，与粒径信息具有较高的一致性。此外，采用了渐进标记策略来提高处理效率和准确性，这将使转移应用能够应用于各种晶体的溶液结晶过程中。提取出各个粒径信息，与粒径信息具有较高的一致性。此外，采用了渐进标记策略来提高处理效率和准确性，这将使转移应用能够应用于各种晶体的溶液结晶过程中。