One-Shot Learning for Robust Material Classification Using Millimeter-Wave Radar System

Wireless classification of different types of objects and materials has great potential in industrial and consumer applications. Different materials have a unique signature of electromagnetic signals that are reflected back to the radar sensor. The two main challenges are to handle instinctive interclass differences and large intraclass variations, as well as sensor variability arising, e.g., from different wafer lots. The limited amount of training data is the limitations for conventional deep learning approaches. We propose to address the issue of material classification in such consumer context using a Siamese network that uses the distance-based similarity metric to be small for same materials and large for different materials. We demonstrate our framework by classifying five variations of four materials using a short-range 60-GHz compact radar sensor achieving an overall accuracy of 99.23%.