Improving remote material classification ability with thermal imagery

Material recognition using optical sensors is a key enabler technology in the field of automation. Nowadays, in the age of deep learning, the challenge shifted from (manual) feature engineering to collecting big data. State of the art recognition approaches are based on deep neural networks employing huge databases. But still, it is difficult to transfer these latest recognition results into the wild – various lighting conditions, a changing image quality, or different and new material classes are challenging complications. Evaluating a larger electromagnetic spectrum is one way to master these challenges. In this study, the infrared (IR) emissivity as a material specific property is investigated regarding its suitability for increasing the material classification reliability. Predictions of a deep learning model are combined with engineered features from IR data. This approach increases the overall accuracy and helps to differentiate between materials that visually appear similar. The solution is verified using real data from the field of automatized disinfection processes.

Description

This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Version

Published version

Access right on openHSU

Metadata only access