Deep learning-based super-resolution reconstruction methods can effectively alleviate the low efficiency of terahertz imaging in detecting defects in polyethylene pipeline fusion joints. However, they are prone to generating artifacts under extremely low-resolution conditions, leading to distorted reconstruction results. To address this, this paper proposes an enhanced super-resolution generative adversarial network that incorporates an efficient multi-scale attention module and employs a vision transformer-based discriminator to achieve super-resolution reconstruction of terahertz images. First, a vision transformer-based discriminator is designed to enhance the ability of the network to discern global image consistency. Subsequently, an efficient multi-scale attention module is incorporated into the generator to improve the network’s focus on the image texture and high-frequency details. Additionally, a composite loss function is constructed to guide the generator in restoring edge information. Compared with five mainstream methods, the proposed approach demonstrates superior performance in PSNR, SSIM, and FLOPs metrics. This method has been proven to effectively enhance the reconstruction quality of terahertz images for hole and crack defects, demonstrating significant potential in the field of defect detection for polyethylene pipeline fusion joints.
sun et al. (Mon,) studied this question.