A novel detail-enhanced wavelet domain feature compensation network for sparse-view X-ray computed laminography

Area of Science:

• Industrial imaging
• Non-destructive testing
• Deep learning applications

Background:

• X-ray Computed Laminography (CL) is vital for non-destructive visualization of flat industrial objects.
• High-quality CL requires numerous projections, leading to lengthy imaging times.
• Reconstructing images from limited projections (sparse-view) degrades quality.

Purpose of the Study:

• Develop a deep learning network for effective sparse-view CL reconstruction.
• Address the trade-off between imaging speed and image quality.
• Enhance the accuracy of CL imaging with reduced data acquisition.

Main Methods:

• An encoder-decoder network architecture was designed, integrating spatial and wavelet domains for feature extraction.
• A detail-enhanced module was incorporated to improve image clarity.
• Hybrid approach combining Swin Transformer and convolution operators for robust feature capture.

Main Results:

• The proposed network achieved superior image quality on solder joint datasets, with PSNR of 37.875 ± 0.908 dB and SSIM of 0.992 ± 0.002.
• Demonstrated high performance using 16-view CL images compared to 1024-view.
• Achieved competitive results on the AAPM dataset, indicating strong generalization.

Conclusions:

• The developed deep learning network is effective for sparse-view CL reconstruction.
• The approach successfully compensates for information loss in reduced-projection imaging.
• The method shows promise for accelerating CL processes without compromising essential diagnostic information.