【摘要】 太赫兹成像技术受辐射源和探测器性能限制，在细节分辨能力、成像速度和噪声抑制等方面仍有进一步改进的空间。本文提出一种基于空间曲线填充的太赫兹图像超分辨算法，采用视觉自注意力（Vision Transformer, ViT）结构主干网络，通过注意力机制进行太赫兹图像特征提取；构建希尔伯特空间曲线，根据特征图按曲线填充的方式进行图像重建，并用轻量化的一维卷积处理重建图像特征，对重建图进行逆变换恢复图像空间结构；最终通过像素重组实现上采样，得到物体轮廓和细节增强的输出图像。实验表明，相较常规ViT结构，本文方法图像峰值信噪比（PSNR）提高0.81 dB，结构相似度（SSIM）提高0.007 4，有效抑制了噪声对图像纹理的影响，获得了分辨能力显著提高的结果图像，证明了太赫兹图像超分辨处理技术的可行性及其恢复图像细节、提高图像质量的能力。更多还原

【Abstract】 The performance of radiation sources and detectors currently limits terahertz imaging technology, which still requires further improvement in terms of detail resolution, imaging speed, and noise suppression. This paper proposes a terahertz image super-resolution algorithm based on spatial curve filling. The ViT(Vision Transformer) structure backbone network is utilized to extract terahertz image features through an attention mechanism. A Hilbert spatial curve is constructed to reconstruct the image according to the feature map using the curve filling method. Lightweight one-dimensional convolution processing is used for reconstructing image features, while inverse transformation of reconstructed maps restores the image’s spatial structure. Finally, pixel reorganization enables up sampling to obtain an output image with enhanced object contour and details. Experimental results show that compared with conventional ViT structures, this proposed method improves Peak Signal-to-Noise Ratio(PSNR) by 0. 81 dB and Structural Similarity Index(SSIM) by 0. 007 4, which effectively inhibits the noise influence on texture and significantly improves the resolution and image quality.更多还原