Terahertz computed tomography (THz CT) demonstrates its advantages in aspects of nonmetallic and nonpolar materials penetration, 3D internal structure visualization, etc. To perform satisfied reconstruction results, it is necessary to obtain complete measurements from many different views. However, this process is time-consuming and we usually obtain incomplete projections for THz CT in practice, which generates artifacts in the final reconstructed images. To address this issue, dictionary learning-based THz CT reconstruction (DLTR) model is proposed in this study. Especially, the image patches are extracted from other state-of-the-art reconstructed images to train the initial dictionary by using the K-SVD algorithm. Then, the dictionary can be adaptively updated during THz CT reconstruction. Finally, the updated dictionary is used for further updating reconstructed images. In order to verify the accuracy and quality of DLTR method, the filtered back-projection (FBP), simultaneous algebraic reconstruction technique (SART), and total variation (TV) reconstruction are chosen as comparisons. The experiment results show that the DLTR method has a good capability for noise suppression and structures preservation.

太赫兹计算机断层扫描 (THz CT) 在非金属和非极性材料穿透、3D 内部结构可视化等方面展示了其优势。要获得满意的重建结果，需要从许多不同的角度获得完整的测量结果。然而，这个过程非常耗时，在实践中我们通常会得到不完整的太赫兹 CT 投影，这会在最终重建图像中产生伪影。为了解决这个问题，本研究提出了基于字典学习的太赫兹 CT 重建（DLTR）模型。特别是，图像块是从其他最先进的重建图像中提取的，以使用 K-SVD 算法训练初始字典。然后，可以在太赫兹 CT 重建期间自适应更新字典。最后，更新后的字典用于进一步更新重建图像。为了验证DLTR方法的准确性和质量，选择滤波反投影（FBP）、同时代数重建技术（SART）和全变差（TV）重建作为比较。实验结果表明，DLTR方法具有良好的噪声抑制和结构保存能力。