The feasibility of applying computer vision sequences to automatically determine the composition of heterogeneous disperse systems, using emulsions as a case study, has been considered. This expands the analytical framework, reduces human factor impact on analysis accuracy and reliability, as well as improves processing speed. During the study, zero-shot segmentation was performed on microscopy images using four different segmenters. The resulting segments were then fitted to circles using a bounding volume (BV) approach. Segmentation effectiveness was evaluated with the Intersection over Union (IoU) metric by comparing results to manually annotated masks provided by an operator. The average IoU values for the applied segmentation models range from 0.64 to 0.68. Applying the BV technique improves agreement with reference masks; specifically, the average IoU fitted to circles reaches approximately 0.75. The overall effectiveness of applying the proposed automatic system in the form of a segmentation and bounding volume sequence was determined by analyzing the emulsion droplet diameter distributions. Comparison of the distributions showed that the data obtained using the automatic system are consistent with the operator's data for fractions larger than 15 px. However, the automatic system underestimates the share of fine fractions, which leads to a systematic shift in the integral assessment. Importantly, it was established that applying the BV method to each individual mask obtained from segmentation is approximately 40–60% faster than analyzing a single combined mask. This analysis of individual masks is also practically more useful in cases involving touching droplets
Kosenko et al. (Fri,) studied this question.