A comparative study of machine learning for particle morphology discrimination using interferometric particle imaging

• Automatic and real-time discrimination of particle morphology in the mixed particle field. • Establishment of a particle images dataset using IPI experimental system and optical transfer matrix theory. • Feature extraction of interferometric defocused particle images using four methods (GM, GLCM, GLDS and GT). • Six machine learning classifiers employed in particle morphology discrimination. • A comparative study of machine learning classifiers based on accuracy, precision, recall and F1-score. The measurement of particles has significant importance in various scientific fields, with the morphology discrimination of particles serving as a prerequisite for accurate measurement. Here, the morphology discrimination of particles is realized by combining interferometric particle imaging (IPI) with machine learning (ML). Particle images obtained from simulation and experiment are processed to extract features based on gray moment, gray level co-occurrence matrix, gray level difference statistics and gabor transform. The extracted feature information is sent into machine learning classifiers: Support Vector Machines, Random Forests, Naïve Bayes, K Nearest Neighbors, Decision Trees and BP Neural Networks. The results show that all classifiers can realize particle morphology discrimination. For this work, the K Nearest Neighbors technique is found to provide the best performance in particle morphology discrimination, with accuracy, precision, recall and F1-score of 97.23%, 95.05%, 100.00% and 97.46%, respectively. The overall evaluation result of KNN is 97.43%.