THz-Net: a multi-scale self-attention residual network for cross-system terahertz image reconstruction

High-resolution reconstruction of terahertz (THz) images is critical for a wide range of imaging applications, such as biomedical diagnosis and non-destructive testing. However, THz images inherently suffer from diffraction-induced blur, low contrast, and loss of fine structural detail. Existing reconstruction methods are limited by insufficient high-frequency feature extraction, lack of multi-scale feature integration, and poor generalization across imaging systems with different optical configurations and noise characteristics. To address these limitations, we propose THz-Net, a multi-scale self-attention residual network for cross-system THz image reconstruction. The network employs an adaptive residual backbone to preserve subtle contrast variations, a multi-scale feature extraction module to capture complementary structural information at different network depths, and a dual-attention module that sequentially applies channel and spatial attention to selectively enhance informative features while suppressing background interference. Experimental results show that THz-Net achieves an average peak signal-to-noise ratio (PSNR) of 37.41dB and an average structural similarity index measure (SSIM) of 0.97, outperforming the best baseline by 3.93dB in PSNR and 0.08 in SSIM, while maintaining consistent reconstruction quality across all three systems.