Identification, classification, and quantification of three physical mechanisms in oil-in-water emulsions using AlexNet with transfer learning

Abstract Physical mechanisms of emulsion are generally observed by microscopy images and subjectively identified or judged by experimenters. However, results are not scientific or convincing due to the lack of specific qualitative or quantitative indicators. To overcome this drawback, AlexNet with transfer learning was employed to automatically identify, classify, and quantify three different physical mechanisms of emulsions. The proposed network achieved good performance with high classification accuracy, and fast training and testing time. Feature visualization of the last fully connected layer represents the common and high-level features of each mechanism, especially the feature image of coalescence, which clearly shows a large droplet is consisting of two or more merged small droplets. Moreover, information entropy calculated the disorder level in feature images of each mechanism, and strongest activations demonstrated the proposed network learns correct features. Therefore, these results contribute to a better understanding of emulsion science from the perspective of deep learning.