Super-resolution reconstruction of terahertz images based on a deep-learning network with a residual channel attention mechanism

Abstract

To date, the existing terahertz super-resolution reconstruction methods based on deep-learning networks have achieved noteworthy success. However, the terahertz image degradation process needs to fully consider the blur and noise of the high-frequency part of the image during the network training process, and cannot be replaced simply by interpolation, which has high complexity. The terahertz degradation model is systematically investigated, and effectively solves the above problems by introducing the remaining channel mechanism into the deep-learning network. On the one hand, an image degradation model suitable for the terahertz imaging process is adopted for the images in the training dataset, which improves the accuracy of network training. On the other hand, the residual channel attention mechanism is introduced to realize the adaptive adjustment of the dependence between network channels, which results in the network being more focused on the restoration of high-frequency information, thereby supporting the extraction of high-frequency edge details in the image. In addition, experimental results demonstrate that this method successfully improves the peak signal-to-noise ratios, and offers clearer edge details and a better overall reconstruction effect. We believe that this work may provide a new possibility to improve the resolution of terahertz images.

J-Net: Improved U-Net for Terahertz Image Super-Resolution

“…Traditional methods, including deconvolution techniques like the Lucy-Richardson algorithm and various interpolation methods, have been instrumental in initial improvements. However, these approaches often fall short in recovering high-frequency details and handling noise variations [1][2][3].…”

Super-Resolution Reconstruction of Terahertz Images Based on Residual Generative Adversarial Network with Enhanced Attention

“…In 2021, Su et al [ 22 ] suggested a novel subspace-and-attention-guided restoration network (SARNet), which adopted attention guidance to fuse spatio-spectral features of amplitude and phase. Yang et al [ 23 ] introduced the remaining channel mechanism and the residual channel attention mechanism to restore the high-frequency information.…”

Image restoration algorithm for terahertz FMCW radar imaging

“…While utilizing only the prior information of the blurred image itself, the total variation [34] and the normalized sparsity measurement [35] blinddeconvolution methods were also used to estimate the PSF and improve the lateral resolution, but the estimation accuracy and restoration capability are limited. In recent years, thanks to the powerful end-to-end mapping and learning capabilities of convolutional neural networks, a variety of deep learning methods [36][37][38][39][40] have also been proposed to implement the THz image super-resolution and have achieved good restoration results. However, when the system parameters (operating frequency, quasi-optical device parameters) are changed, the network needs to be retrained.…”