X-ray tomography is a nondestructive technique that visualizes interior features in solid objects. To achieve a given resolution, a sufficient number of projection images over a cycle is required for 3D reconstruction based on the Crowther criterion. However, practical limitations such as geometrical constraints, data acquisition time, and low dose requirements often prohibit the acquisition of the full dataset, only allowing a limited angular range. The unsampled angles lead to the 'missing edge' problem in tomography, which introduces strong artifacts in reconstruction. To tackle this challenge, we propose an approach that integrates a Convolutional Neural Network (CNN) as a regularizer into an iterative solving engine. It combines perceptual prior knowledge about the sample with the physical model to produce an artifact-free solution. Our approach demonstrates excellent results with an experimental dataset with a missing edge of over 90 degrees.
|