- The paper introduces RCL-Mamba, a framework that integrates projection domain correction and image domain refinement to counteract measurement degradations in RCL.
- It employs a dual-branch SSM-CNN module with dynamic gated fusion to balance global blur modeling and high-frequency detail preservation, achieving up to +2.98 dB PSNR improvement over competitors.
- The method enables reliable restoration from as few as 64 views, reducing scanning time by ~8x while maintaining critical measurement fidelity in industrial PCB inspection.
Measurement-oriented Restoration in Sparse-view RCL: The RCL-Mamba Framework
Context and Motivation
Rotational Scanning Computed Laminography (RCL) has become indispensable in nondestructive testing (NDT) for large planar structures, especially in high-throughput industrial environments like PCB inspection. Industrial requirements for rapid throughput necessitate continuous sparse-view data acquisition, which introduces intrinsically coupled cross-domain degradations: severe rotational blur in the projection domain (from angular integration during exposure) and sparse-view artifacts in the reconstructed image domain. Prior CNN-based and Transformer-based restoration schemes show evident performance bottlenecks: CNNs are hampered by local receptive field constraints and a stationarity assumption, while vision Transformers, although more global, incur high computational cost and can propagate artifacts due to lack of inductive bias about RCL imaging physics. Recent visual state space models (VSSMs) such as Mamba address long-range dependencies more efficiently, but most are tailored to natural images and ignore cross-domain physical coupling fundamental to RCL.
Technical Contributions
RCL-Mamba introduces a measurement-oriented dual-domain restoration strategy explicitly matched to the physical imaging and degradation sequence in RCL. The architecture is a two-stage cascaded system:
- Projection Domain Correction (P-Net): Targets and corrects rotational blur at the projection level, addressing the dominant front-end degradation source before reconstruction.
- Image Domain Refinement (I-Net): Suppresses residual artifacts and local distortions arising from sparse-view sampling, operating on reconstructed tomographic slices.
To balance the need for both global degradation modeling (large-scale, spatially variant blur) and preservation of high-frequency microstructures, each network employs a Mamba-CNN dual-branch module with dynamic gated fusion. The Mamba branch applies a selective scan mechanism (S6) to efficiently model global, highly anisotropic degradation trajectories, leveraging the state space model's linear complexity. The CNN branch simultaneously captures high-frequency edge and texture information. The fusion mechanism adaptively weights the contribution of both branches in a spatially-aware manner.
Loss functions are formulated to jointly optimize pixel-domain, frequency-domain, and structural edge fidelity, supporting robust recovery of microstructural measurement features critical in NDT.
Empirical Evaluation and Analysis
RCL-Mamba is validated on both simulated and real NDT datasets generated from industrial-grade PCB phantoms and active RCL inspection systems. The method exhibits:
- Superior quantitative results: On simulated data, RCL-Mamba achieves PSNR 38.46 dB, SSIM 0.9862, and GMSD 0.0099, outperforming the second-best competitor (MLWNet-B) by +2.98 dB PSNR. On real data, the method achieves PSNR 32.47 dB, SSIM 0.9543, and the lowest GMSD, demonstrating robust generalization and fidelity under practical noise and artifact scenarios.
- Improved structural measurement consistency: Line-profile-based analysis shows that RCL-Mamba minimizes RMSE with respect to the 512-view reference for both via/pad and trace-edge profiles (mean RMSE 0.0341), confirming the preservation of critical geometrical boundaries often degraded by single-domain or non-physically-aligned approaches.
- Dramatically increased inspection efficiency: The dual-domain strategy enables reliable restoration from as few as 64 sparse views (versus 512 in the reference setting), reducing effective scanning time by ~8x, without compromise in measurement-oriented slice fidelity.
Ablation studies affirm both the necessity of dual-domain decoupling—where error propagation is greatly reduced by correcting the dominant rotational blur prior to reconstruction—and the effectiveness of the dual-branch dynamic fusion for global/local feature balance.
RCL-Mamba further achieves this with a model footprint of only 17.2M parameters (substantially smaller than leading CNN competitors), with moderate FLOPs, evidencing a favorable trade-off between accuracy and deployment cost for industrial scenarios.
Implications and Future Directions
The RCL-Mamba architecture instantiates a paradigm where network design is tightly coupled to the physics of measurement and degradation. This measurement-oriented, cross-domain-corrective approach aligns restoration with the true underlying image formation chain, unlike standard single-domain or black-box models. Practically, it enables a step change in industrial RCL inspection throughput, unlocking reliable high-speed NDT in settings previously bottlenecked by reconstruction quality at low acquisition rates. The dual-branch SSM-CNN modules offer a new standard for tasks requiring both long-range global blur modeling and precise edge fidelity under severe, spatially-varying artifacts.
Theoretically, RCL-Mamba exposes avenues for further work in integrating stronger geometric and physics-informed priors into VSSMs (e.g., polar coordinate transforms for rotationally-varying blur trajectories), moving beyond generic 2D scanning strategies. Closing the loop between physical measurement, degradation path, and neural architecture will be increasingly essential in ambitious, measurement-critical vision applications.
Unresolved challenges include further quantification of absolute measurement uncertainty, especially in sub-millimeter PCB feature regimes, and generalization of the approach to arbitrary or unknown imaging geometries, as well as incorporation of real-time constraints for in-line deployment.
Conclusion
RCL-Mamba demonstrates a principled, physically-aligned solution to the coupled restoration problem in rotational sparse-view RCL for industrial inspection. By integrating dual-domain cascaded correction with a SSM-CNN dual-branch module, it achieves substantial gains in both measurement fidelity and throughput efficiency compared to prior single-domain or general-purpose VSSM baselines. These advances both highlight the necessity of physics-aware design in measurement-oriented imaging AI and establish a foundation for further exploration of state space models in a broader set of constrained vision tasks.
