<p>Computer laminography (CL) system shows great potential as a non-invasive method for visualizing the internal structures of objects with a relatively flat shape. However, the lack of projection data from certain angles in the CL system leads to severe streak artifacts when conventional reconstruction algorithms are used. To address this issue, we propose a deep learning approach for image post-processing for fast CL system imaging. Since these systems do not conform to the standard geometry of the FDK algorithm, the reconstruction process involves reprojecting original projections (RP) to those of the standard geometry and applying the FDK algorithm to obtain the reconstruction results (RP + FDK). To reduce limited-angle artifacts of the degraded images, we then employ MIARNet, which utilizes a multi-scale feature fusion block to exchange information across different scales. Moreover, a frequency selection loss function is introduced to suppress the unique structures of limited-angle artifacts in the frequency domain of the degraded images. Simulations and real data experiments demonstrate that our reconstruction method significantly reduces limited-angle artifacts while preserving high-resolution details.</p>