LaminoDiff: Artifact-Free Computed Laminography in Non-Destructive Testing via Diffusion Model

TL;DR

LaminoDiff tackles artifact-rich Computed Laminography by recasting reconstruction as a conditional diffusion process guided by a physics-informed CT-CL fusion prior. A dual-scale fusion target $x_{fus}$, derived from CT and CL data, provides a high-fidelity supervisory signal that narrows the sim-to-real domain gap while leveraging the CL observation $x_{fdk}$. The method uses a spatially coupled denoiser with time-aware embeddings, conditioned on $x_{fdk}$ and $x_{fus}$, and performs deterministic DDIM reconstruction with classifier-free guidance to enforce data-consistency and structural fidelity. Across simulated and real PCB datasets, LaminoDiff achieves superior artifact suppression and edge preservation, enabling more reliable automated defect recognition in industrial inspection. A remaining challenge is axis anisotropy due to the missing cone; future work includes pseudo-3D spatial constraints or interlayer attention to further isotropize resolution.

Abstract

Computed Laminography (CL) is a key non-destructive testing technology for the visualization of internal structures in large planar objects. The inherent scanning geometry of CL inevitably results in inter-layer aliasing artifacts, limiting its practical application, particularly in electronic component inspection. While deep learning (DL) provides a powerful paradigm for artifact removal, its effectiveness is often limited by the domain gap between synthetic data and real-world data. In this work, we present LaminoDiff, a framework to integrate a diffusion model with a high-fidelity prior representation to bridge the domain gap in CL imaging. This prior, generated via a dual-modal CT-CL fusion strategy, is integrated into the proposed network as a conditional constraint. This integration ensures high-precision preservation of circuit structures and geometric fidelity while suppressing artifacts. Extensive experiments on both simulated and real PCB datasets demonstrate that LaminoDiff achieves high-fidelity reconstruction with competitive performance in artifact suppression and detail recovery. More importantly, the results facilitate reliable automated defect recognition.

Figures (13)

• Figure 1: Scanning geometry of CT and CL. (a) CT geometry with orthogonal rotation. (b) A schematic diagram of the spectrum in the Fourier domain. (c) CL geometry featuring an inclined rotation axis. (d) The commercial CL inspection system at the Institute of High Energy Physics, Chinese Academy of Sciences (Beijing).
• Figure 2: Workflow of the CL-based inspection system. The process typically involves (a) data acquisition using the CL system, (b) projection data collection, (c) image reconstruction, and (d) final defect detection and reporting.
• Figure 3: Schematic of the dual-scale iterative fusion framework. The process begins by calibrating the intensity of the high-resolution CL data with the complete CT data. These inputs are then fed into the iterative fusion strategy. Through an alternating update scheme involving both a coarse grid ($x_G$) and a fine grid ($x_H$), the algorithm synergizes the spatial precision of CL with the completeness of CT, yielding a high-fidelity fusion target $x_{fus}$.
• Figure 4: Comparison of different targets. The figure evaluates the signal features of four label strategies across spatial (top row), histogram (middle row), and spectral (bottom row) domains. (a) The original reconstruction results contain aliasing artifacts. (b) Numerical phantom with statistically binary. (c) The pseudo-labels generated by CycleGAN-CL. (d) The pseudo-labels generated by CycleGAN-CT. (e) The proposed CT-CL fusion label.
• Figure 5: Schematic illustration of the proposed diffusion-based reconstruction framework. (a) Forward Diffusion Process: Gaussian noise is progressively added to the clean fusion label $x_{fus}$ to generate the noisy state $x_t$. (b) DDIM Inference: The artifact-free image is iteratively reconstructed from noise via DDIM, conditioned on the CL $x_{fdk}$. (c) Training Architecture: A U-Net with skip connections processes the concatenated inputs ($x_t$, $x_{fdk}$, $x_{fus}$) using ResBlocks and Attention mechanisms to extract multi-scale features. (d) Optimization Objective: The model is trained to predict the added noise $\epsilon_\theta$ by minimizing the loss.