An Integrated Histogram-Based Vision and Machine-Learning Classification Model for Industrial Emulsion Processing

Existing techniques in emulsion quality evaluation are found to be highly subjective, time-consuming, and prone to overprocessing. Other conventional droplet analysis techniques such as laser diffraction, which require dilution of samples, introduce an additional complexity to industrial processes. The possibility of developing a fully automated technique for droplet characterization during emulsification holds remarkable potential for overcoming the existing challenges. In this article, a histogram-based image segmentation technique detects droplets from emulsion micrographs. The evolution of droplet characteristics and their significance are studied by performing statistical analysis, and the significant characteristics are selected. The principal component analysis is applied to obtain a reduced set of uncorrelated components from the selected characteristics. The linear discriminant analysis classifies the micrographs into a set of quality categories called target, acceptable, marginal, and unacceptable. The model accuracy is validated using stratified five-fold cross-validation and is successful in classifying the micrographs obtained from two different manufacturing facilities with high accuracy up to 100%. The histogram-based technique is successful in detecting smaller droplets than previously reflected in the literature. The current approach is fully automated and is implemented as a soft-sensor, which supports its real-time deployment into an industrial environment. The entire approach has promising potential in the in-line prediction of emulsion quality leading to more efficient and sustainable manufacturing.