Mask-RSCN: An Improved Multiclass Object Detection Algorithm for Crystallization Process

Accurate monitoring and information extraction of crystals remain a significant challenge and research focus in industrial crystallization processes. In this article, a new mask-residual-spatial-channel network (Mask-RSCN) algorithm is proposed for in situ image detection of multiclass objects with higher accuracy and efficiency, which was improved based on the mask-region-based convolutional neural network (Mask-RCNN) algorithm by integrating three modules into the feature pyramid network (FPN). An additional residual module (ARM) was introduced into Mask-RCNN before feature fusion to improve the classification accuracy of multiscale crystals, while the spatial refinement module (SRM) and channel refinement module (CRM) were used to improve the structure of input and output of feature fusion, respectively, so as to improve the extraction accuracy of crystal positioning and classification. Three datasets were employed to test the performance of the proposed approach and compared with the Mask-RCNN algorithms. Results demonstrated that the detection accuracy of Mask-RSCN was improved by more than 30% with less loss. In addition, other convolutional neural networks (CNNs) were further to verify the advantage of the proposed method with strong generalization ability.