High-resolution parallel-translation CT reconstruction based on a patch-based adaptive-regularization-guided diffusion model.
Authors
Affiliations (1)
Affiliations (1)
- School of Information and Communication Engineering, Dalian University of Technology, Dalian 116024, China.
Abstract
Computed tomography (CT) is an effective technique for industrial non-destructive testing (NDT). However, conventional rotational CT is difficult to apply to the scanning and reconstruction of large plate-like objects, while its mechanical system is also relatively complex. In response to these limitations, parallel translation CT (PTCT) has been introduced. Nevertheless, due to its limited scanning range, PTCT acquires incomplete projection data, which leads to aliasing artifacts and makes accurate reconstruction challenging. To address these issues and achieve high-resolution PTCT reconstruction, this paper proposes a patch-based adaptive-regularization-guided diffusion model. Specifically, an initial estimate obtained by the Simultaneous Iterative Reconstruction Technique (SIRT) is first employed, introducing structural prior information into the reconstruction process. Based on this, a split-and-reassemble denoising module (SRDM) is designed to perform patch-wise reconstruction, where positional encoding is incorporated as conditional guidance to enhance the representation of local structural details. Furthermore, an adaptive-regularization-based correction module (ARCM) is introduced, in which seam-based and mask-based regularization terms are adaptively adjusted according to the denoising stage to support global structural recovery and improve local structural continuity. To evaluate the proposed method, a PCB dataset containing samples with defect structures is constructed. The results demonstrate that the proposed method achieves reliable performance in global reconstruction while effectively recovering fine structural details within defect regions.