Complex Valued Neural Networks for Terahertz Computational Imaging

The vast majority of deep learning algorithms are based on real valued operations, thereby neglecting the physical characteristics inherent in complex data. This paper presents a complex valued neural network for terahertz computational imaging. Based on the mathematical properties of complex valued convolution operations, the operations of convolutional layers and fully connected layer in the neural network structure are reconstructed. The amplitude and phase information of the received signal are taken into account, which improves the ability of the algorithm to reconstruct the image. The image restoration capability of the complex valued neural network under low sampling rates, as well as its resistance to noise interference, has been validated through simulation experiments.