Convolutional Neural Networks for Chipless RFID Classification in the Time-Frequency Domain

This paper presents a novel approach to accurately classify chipless RFID measurements in the time-frequency domain via Convolutional Neural Networks (CNN). Magnitude and Phase data of S-11 measurements, ranging in distance 50-140 cm from the emitter, are transformed to time-domain signals using the Inverse Fast Fourier Transform (IFFT). A matrix representation of the time-domain signal is then obtained using the Continuous Wavelet Transform (CWT) and used to fit a CNN model. The proposed preprocessing pipeline and the designed CNN architecture achieve up to 86.75% accuracy when used to classify 16 tags, outperforming by up to 10.63% a previously proposed scheme based on magnitude data and Logistic Regression.