Unsupervised aliasing artifact suppression for CBCL via adjacent slice feature fusion GAN

Cone-beam Computed Laminography (CBCL) is a core technology for the non-destructive testing of large plate-shaped samples, yet the missing projection information caused by the inclination angle of its rotation axis leads to aliasing artifacts in reconstructed images and severely degrades the imaging quality. Existing supervised deep learning methods rely on paired clear images and artifact-contaminated images, which are difficult to acquire in practical scenarios. To address this issue, we propose an unsupervised aliasing artifact suppression network based on adjacent slice feature fusion (ASFF-GAN) and design an Adjacent Slice Feature Fused Dual-Input Generator (ASFF-DIG). By introducing adjacent slice information to enhance the network's capability for image structure reconstruction and combining with the CycleGAN framework, unsupervised training is realized. We construct a real dataset with 50×50 cm large-size PCB samples as experimental subjects. Experimental results demonstrate that ASFF-GAN can effectively suppress aliasing artifacts in both complex and smooth regions while fully preserving the horizontal structural information of the samples. Eliminating the need for paired training data, this network provides an efficient solution for aliasing artifact suppression in CBCL imaging and has practical application value in the field of industrial non-destructive testing.