Mask R-CNN based droplet detection in liquid–liquid systems, Part 2: Methodology for determining training and image processing parameter values improving droplet detection accuracy

Abstract

Multiphase systems occur in many technical processes, e.g., crystallization, gas–liquid reaction, and liquid–liquid extraction. Operating such processes requires knowledge of fluid dynamics, mass transfer, and reaction kinetics that depend on the phase interface. The particle size distribution quantifies the phase interface of multiphase systems. Multiphase systems can contain solid particles, gas bubbles, or liquid droplets. Our previous work presented an automated and publicly accessible image processing method to determine droplet size distributions (DSD) in liquid–liquid systems based on Mask Region-Based Convolutional Neural Networks (MRCNN). In this work, we developed a methodology to determine training and image processing parameter values that improve the droplet detection accuracy of our MRCNN implementation with a low computational effort. Our methodology combines a sensitivity analysis to reduce the dimensions of training parameters of interest with a systematic final model selection approach based on characteristic training and image processing rating metrics. Applying our methodology to a database of four different image qualities results in a highly improved droplet detection accuracy over the entire droplet diameter range of a DSD. In comparison to the MRCNN models trained on the default hyperparameter values, the average relative error in characteristic DSD properties: minimum diameter, maximum diameter, and the number of detected droplets could be reduced from 6% to 3%, from 7% to 2%, and from 7% to 2%, respectively, keeping the relative error in the Sauter mean diameter constant at 3%. The presented methodology can be universally applied to individual image qualities improving the droplet detection accuracy of our MRCNN implementation. Thus, combining our methodology with the MRCNN implementation increases the reliability of process analysis and process control in liquid–liquid systems. Moreover, our MRCNN implementation and methodology can likely be easily transferred to several applications in which particle detection plays a significant role, e.g., crystallization (solid particles) and gas–liquid reaction (bubbles).