Artifact reduction in rotational computed laminography using a deep learning method

Computed laminography (CL) is widely used in imaging plate-like object. Due to lacking projection data along non-thickness direction, images reconstructed from CL contain severe interlayer aliasing artifacts. These artifacts can greatly affect later object identification and information extraction from the CL images, restricting their utility value. To reduce aliasing artifact in CL, we develop a deep learning method to post-process the FDKreconstructed image. Firstly, we analyze the characteristics of aliasing artifacts in CL images. Based on that, a modified U-Net convolutional neural network (CNN), which takes 2.5D radial slice as input, is proposed. Then, the effectiveness of the proposed method is tested and compared with other strategies, including the methods using 2D (z direction, x direction, and radical direction) slice and 2.5D (z direction, x direction) slice as input. Experimental results on ball grid array (BGA) specimens shows that the proposed method give the best performance in the CL aliasing artifact reduction in all comparison strategies.