Abstract
We present supervised terahertz deep learning models for high-precision terahertz tomography. To investigate the performance of terahertz deep learning models, comprehensive characterization of kernel size in first convolution layers is further studied. By utilizing the length of beam diameter, the optimized kernel size can be designed to deliver much spatially accurate images, which achieves 2.5% on mean square error (MSE), exhibiting 46.8% improvement on MSE than other kernel sizes.