Deep Learning-Based Method for Removing Inter-layer Aliasing Artifacts in Computed Laminography Reconstruction

Computed laminography (CL)—featuring a tilted scanning configuration—mitigates the constraints of computerized tomography on X-ray energy and scanning space when inspecting plate-like objects. However, this design compromises the completeness of projection data, thus causing interlayer aliasing artifacts in reconstructed images. Hence, this study proposes an image post-processing approach. Specifically, a U-shaped neural network model integrating Swin Transformer blocks and attention mechanisms is developed to refine low-quality images corrupted by aliasing artifacts. An anisotropic composite loss function is derived to enhance the restoration of edge details in images. Finally, a feature-fusion technique is employed to correct the network’s output. Experiments conducted on simulated printed circuit board datasets demonstrate that the proposed method improves the PSNR and SSIM of SART-reconstructed results by 144.73% and 68.10%, respectively. It effectively eliminates aliasing artifacts and recovers the correct structural information of objects.