Terahertz imaging presents immense potential across many fields but the affordability of multiple-pixel imaging equipment remains a challenge for many researchers. To address this, the adoption of single-pixel imaging emerges as a lower-cost option, however, the data acquisition process necessary for reconstructing images is time-intensive. Compressive Sensing, which allows for generation of images using a reduced number of measurements than Nyquist's theorem demands, presents a promising solution but long processing times are still issue particularly large-sized images. Our proposed solution to this issue involves using caustic lens effect induced by perturbations in a ripple tank as a sampling mask. The dynamic nature of the ripple tank introduces randomness into the sampling process and this reduces measurement time by exploiting the inherent sparsity of THz band signals. This work employed Convolutional Neural Network to perform target classification based on the distinct signal patterns acquired through the caustic lens mask. The proposed classifier achieved 99.22% accuracy rate in distinguishing targets shaped like Latin letters. The controlled randomness introduced by the caustic lens mask is believed to play a crucial role in achieving this high accuracy by mitigating overfitting, a common challenge in machine learning.
Read full abstract