Nearfield for Classification of Mouth–Neck Activities Based on Wavelet Filtering and Gabor Transformation Using DCNN

This article proposes a novel method for detecting human mouth-neck activities using the nearfield effect of a single antenna used as the noninvasive sensor capable of capturing very complex and subtle variations of the surroundings. The time-varying signatures of reflection coefficient (
${S}_{{11}}$
) due to perturbed nearfield radiation reveal minute distinct characteristics gathered from the Flexible-RT/Duroid (F-RTD) antenna, which can pinpoint the distinct qualities and coexisting changes in radiation of these actions for various mouth-neck motions. The F-RTD antenna demonstrates good biocompatibility to electromagnetic absorption. The wavelet-based denoising of perturbed
${S}_{{11}}$
signatures is obtained using decomposition, elimination of coefficients, and reconstruction process. A deep learning (DL) model is used to extract deep features from 2-D time-frequency (TF) images generated by a Gaussian windowed Gabor wavelet transform (GWT) of the denoised signals. Two more pretrained benchmark models, namely, Squeeze-Net and Alex-Net, are used, and it is found that the proposed DL model produces the best outcome in every case. The antenna is placed in two distinct locations, on the neck and near the mouth. The neck mount is shown to be effective localization for detecting six different motions. The classification performance is analyzed
${S}_{{11}}$
magnitude and phase, as well as several sliding time windows. At the optimal sliding frame of 2.5 s, the accuracy (ACC),
${F}1$
score, sensitivity (SEN), and precision (PRC) metrics for neck mount are found to be 94.68%, 94.7%, 94.68%, and 94.72%, respectively. Similarly, the corresponding metrics for mouth mount are 90.88%, 90.45%, 90%, and 90.88%.