Lightweight mask R-CNN for instance segmentation and particle physical property analysis in multiphase flow

Abstract

A lightweight Mask R-CNN instance segmentation model was developed here to analyze particle size and shape accurately and quickly. Firstly, a hybrid Depthwise Dilated Convolutional Network (DDNet) is proposed, and the feature pyramid layers and the shared convolutional layers of the region proposal network are simplified, reducing the model complexity while ensuring robust feature extraction capabilities. Then, segmentation accuracy is significantly improved without sacrificing computational speed and performance by introducing the Dice loss function and clustering algorithm. Experimental results show that the model parameters are significantly reduced by 49.46%, and the segmentation speed increases from 2.15 FPS (frames per second) to 5.88 FPS. Meanwhile, the segmentation accuracy (AP50) increased from 90.56% to 91.21%. In addition, it was proven that the particle size distribution and shape could be analyzed accurately and rapidly with the proposed model, providing essential information for multiphase flow process optimization and equipment design in industrial applications.