Deep Learning for Range-Doppler Map Single Frame Classifications of Cooking Processes

This paper proposes a Deep Learning approach for microwave frequency based classification tasks using single frame Range-Doppler maps. The Range-Doppler maps are recorded with a 77 GHz chirp-sequence radar sensor. The proposed networks are verified with an application to detect states like boiling in cooking processes. The network achieves an accuracy of 99.17% over six classes while being lightweight and fast. After training, the trained networks are analyzed with a technique that extracts the learned patterns of the network. The effect of pooling layers in convolutional neural networks is discussed due to the loss of detailed information in Range-Doppler maps.