ASFF-GAN: an unsupervised training network for aliasing artifact reduction in cone-beam computed laminography via...

Cone-beam computed laminography (CBCL) is a core technology for the non-destructive testing (NDT) of large plate-shaped objects such as PCBs, BGAs, and ICs. However, the inclination angle of the rotation axis in the geometry of its imaging system leads to missing projection information, which causes aliasing artifacts in the reconstructed images and severely degrades the imaging quality. In existing solutions, supervised deep learning networks require paired clear images and images with aliasing artifacts, which are extremely difficult to obtain in practical scenarios. To address this challenge, this paper proposes an unsupervised aliasing artifact suppression network (ASFF-GAN) based on CycleGAN, aiming to break through the data constraints of supervised training. Targeting the inter-layer information mixing characteristic of aliasing artifacts in CBCL, this paper designs a dual-input generator with adjacent slice feature fusion (ASFF-DIG) as the core of the network. By introducing adjacent slice information of the slice to be processed through a dual-input mode and enhancing the network's perception and reconstruction capabilities via feature fusion, effective suppression of aliasing artifacts is achieved. In this paper, a real training dataset is constructed using a CL imaging system, with large-sized PCB samples as the experimental subjects. Experimental results demonstrate that ASFF-GAN can effectively suppress aliasing artifacts in both complex structural regions and smooth regions, while fully preserving the horizontal structural information of the samples. Furthermore, ablation experiments verify that the adjacent slice information introduced by the dual-input structure is the key factor for improving the aliasing artifact suppression performance. This network eliminates the need for paired training data, demonstrates feasibility in practical imaging applications, and thus provides a novel approach to addressing the issue of aliasing artifact suppression in CBCL imaging.