Image-Based Classification of Bolus Phases in Robotic Mastication Using Hybrid Machine Learning Analysis

Non-invasive, continuous monitoring of food bolus transformation during robotic mastication remains a significant challenge, as existing methods rely on physical extraction of the bolus or single-modality image features that lack robustness across the continuous and non-linear nature of food breakdown. This paper investigates whether hybrid machine-extracted image features can reliably classify chewing phases during in vitro robotic mastication without physical intervention. A hybrid feature-extraction framework is proposed that integrates deep convolutional features from VGG16 with classical texture descriptors, Grey-Level Co-occurrence Matrix (GLCM), image lightness, and variance, applied to a dataset of 1024 segmented chewed biscuit bolus images captured across two chewing trajectories: vertical motion (T1) and combined vertical and lateral motion (T13). Supervised classifiers, Support Vector Machines, Linear Discriminant Analysis (LDA), and Random Forests were evaluated using trial-level grouped cross-validation. Using a 3-phase schema conceptually aligned with the Hutchings and Lillford Mouth Process Model, LDA achieved accuracies of 75.9% for T1 and 82.3% (T13), with the hybrid feature set outperforming the ablation study. Improved separability with T13 demonstrates that shear effect produces more distinctive bolus image features. The unchewed class was consistently well separated, while structured misclassification patterns occur mainly between adjacent phases highlight the challenging nature of the study. These results establish a non-contact, vision-based measurement framework for classifying bolus evolution as a proof-of-concept foundation for robotic chewing systems.