Haptic Material Analysis and Classification Inspired by Human Exploratory Procedures.

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Accurate characterization of object material properties is crucial for various applications.
• Existing methods often lack comprehensive data or efficient acquisition techniques.
• Understanding material properties aids in realistic digital content creation and physical simulations.

Purpose of the Study:

• To introduce a framework for acquiring and parametrizing object material properties.
• To develop a novel set of mathematical features for material classification.
• To evaluate the performance of machine learning models in material identification.

Main Methods:

• An acquisition device, Texplorer2, was used to scan 184 material classes.
• Materials were labeled using biological, chemical, and geological conventions.
• Novel mathematical features were engineered and used with machine learning classifiers, including random forest and deep neural networks.

Main Results:

• The proposed multi-modal features achieved 90.2% ± 1.2% classification accuracy with a random forest classifier.
• A deep neural network achieved 90.7% ± 1.0% accuracy on surface images.
• The deep neural network exhibited more critical misclassifications within the proposed taxonomy compared to the feature-based approach.

Conclusions:

• The developed framework and novel features provide an effective method for material property acquisition and classification.
• Machine learning techniques, particularly random forest with engineered features, show strong performance in material identification.
• While deep learning shows comparable accuracy, the feature-based approach offers better interpretability and fewer critical errors for the defined taxonomy.