Adaptive rectangular convolution and wavelet attention-based GAN for high-fidelity terahertz image super-resolution

Terahertz (THz) imaging, characterized by its non-ionizing radiation and spectral fingerprinting capabilities, has become a pivotal tool for non-destructive testing (NDT) and security inspection applications. However, due to the diffraction limit inherent to long-wavelength radiation and the associated constraints of low-numerical-aperture optics, THz images suffer from low spatial resolution and blurred edge features. To address these challenges, we propose a deep-learning framework, namely, ARWA-GAN, for high-fidelity super-resolution (SR) reconstruction of THz images. Built upon generative adversarial network (GAN), ARWA-GAN introduces two key components: an Adaptive Rectangular Convolution (ARConv) module that autonomously adjusts convolution kernels and receptive fields to extract spatial features more precisely, and a Wavelet Attention (WA) module that enhances high-frequency features through discrete wavelet transform. Experimental results on THz security-screening datasets demonstrate that the proposed ARWA-GAN model achieves superior peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) compared to existing SR algorithms, while also producing reconstructed images with sharper edges, richer texture details, and fewer artifacts in terms of visual perception. These results validate ARWA-GAN as an efficient image enhancement solution for THz imaging in practical scenarios such as NDT and security inspection. © 2026 Elsevier Ltd.