Isurie Akarawita: Machine Vision for Classifying Bolus Formation in Robotic Chewing

A total of 1,024 bolus images were collected from two chewing trajectories: vertical compression (T1) and combined vertical–shear motion (T13). Images were segmented using SAM2, and deep VGG16 features were fused with classical texture descriptors, including GLCM features, mean lightness, and intensity variance.

Four chew-count-defined stages: Unchewed, Initial, Mid, and Final were classified using LDA, SVM, and Random Forest models with trial-grouped cross-validation. LDA achieved the highest accuracy, reaching 65% for T1 and 68% for T13. The shear-assisted T13 trajectory provided improved stage separability, while most classification errors occurred between neighbouring chewing stages.