TL;DR
This paper introduces a novel 2D Directional Filtered Backprojection (DFBP) method for cone-beam rotary computed laminography (CL) that significantly reduces superimposition artifacts. By deriving a simplified Point Spread Function (PSF) through clever weighting factors and transferring 3D image-space filtering into the 2D projection domain, the authors achieve sharper reconstructions and better defect detection in high-aspect-ratio objects like PCBs.
What Problem Does This Paper Address?
In the world of non-destructive testing, Computed Tomography (CT) is king, but it has a weakness: flat, wide objects with high aspect ratios (like printed circuit boards or battery packs). X-rays struggle to penetrate these objects along their long axis, leading to poor data.
Computed Laminography (CL) solves this by tilting the system so the X-ray beam always hits the object at an angle. However, this tilted geometry introduces a new headache: "superimposition artifacts." Conventional reconstruction algorithms (like FBP or FDK) were designed for standard CT and rely on 1D row-wise filtering. In the complex, nonlinear geometry of rotary CL, this 1D approach fails to account for how information from different layers smears together, resulting in blurry edges, low contrast, and "ghosting" from adjacent slices.
Methods
The authors propose a PSF-Informed 2D Directional Filtered Backprojection (DFBP) method. Their approach follows three main logical steps:
- Exact Geometry Modeling: They first derive the exact relationship between a 3D voxel in the image space and its 2D projection on the detector. This is crucial because, unlike parallel-beam systems, the mapping in cone-beam rotary CL is nonlinear.
- PSF Simplification via Weighting: To make the math tractable, they introduce two specific weighting factors into the backprojection process. This "linearizes" the Point Spread Function (PSF), turning a messy 3D problem into something that looks more like a standard convolution.
- Efficient 2D Filtering: Filtering in 3D image space is computationally ruinous. The authors develop an approximate method to transfer this 3D filtering operation into the 2D projection space. By linearizing the coordinate relationship in Fourier space (under the assumption that the object is relatively thin), they derive a 2D directional filter that integrates von Hann and sinc windows to suppress noise and sharpen boundaries.
Datasets
The method was evaluated using real-world projection data from a physical CT scanning system:
- PCB 1 & PCB 2: Two actual printed circuit boards were scanned. These are classic "high-aspect-ratio" objects where internal details like solder joints and micro-cracks are critical for quality assessment.
- System Specs: The data was acquired with a source-to-rotation center distance of 16.12mm and a tilted angle of 45°. The detector array was 1536 × 1536 pixels.
Key Insights and Ideas
- Moving Beyond 1D: The core "aha" moment is that 1D filtering is fundamentally insufficient for tilted rotary geometries. Information is smeared across the detector in a way that requires a 2D directional approach to "un-smear" correctly.
- The Power of Weighting: By applying weights during backprojection, the authors essentially "pre-correct" for the geometric distortions of the cone beam, allowing the subsequent filtering to be much simpler and more effective.
- Fourier Linearization: The transition from 3D image filtering to 2D projection filtering is the paper's primary engineering feat. It provides the image quality of expensive iterative methods with the speed of analytical ones.
Key Results
The proposed DFBP method was compared against standard CL-FBP and CL-FDK algorithms.
- Visual Clarity: Reconstructions showed significantly sharper boundaries and higher contrast. Fine circuit patterns that were "washed out" in standard methods became clearly defined.
- Defect Detection: The method excelled at identifying subtle internal defects, such as trapped air bubbles in solder joints and electronic components, which were either distorted or invisible in the baseline reconstructions.
- Artifact Suppression: Superimposition artifacts (the "ghostly" shadows of other layers) were noticeably reduced, particularly in the central slices of the volume.
Limitations and Future Work
The primary limitation is the "thin-plate" approximation ($|z| \ll SO \cos \tau$). The math assumes the object is relatively thin and centered. As you move further away from the central reconstruction plane (increasing $|z|$), the approximation starts to break down, and superimposition artifacts begin to creep back in.
The authors suggest that future work will focus on designing even more tailored weighting functions and advanced filtering techniques to mitigate these artifacts in thicker objects or slices further from the center.
Summary updated