Improving remote material classification ability with thermal imagery

Area of Science:

• Automation and Robotics
• Sensor Technology
• Machine Learning

Background:

• Material recognition is crucial for automation, with deep learning dominating current approaches.
• Deep learning models require large datasets but struggle with real-world variations like lighting and image quality.
• Existing methods face challenges in differentiating visually similar materials.

Purpose of the Study:

• To investigate the use of infrared (IR) emissivity for improving material classification reliability.
• To combine deep learning predictions with engineered features from IR data.
• To enhance the robustness of material recognition systems in diverse environments.

Main Methods:

• Utilized deep learning models for initial material recognition.
• Incorporated engineered features derived from infrared (IR) emissivity data.
• Evaluated the combined approach using real-world data from automated disinfection processes.

Main Results:

• The integrated approach significantly increased overall material classification accuracy.
• The method demonstrated improved differentiation between visually similar materials.
• Successful verification was achieved using field data from automatized disinfection.

Conclusions:

• Infrared (IR) emissivity is a valuable material-specific property for enhancing recognition reliability.
• Combining deep learning with IR data overcomes limitations of optical sensors in variable conditions.
• The proposed method offers a practical solution for robust material recognition in automation.