Fast Neural Network Based X-Ray Tomography of Fruit on a Conveyor Belt

Inline computed tomography (CT) based food inspection requires a fast image reconstruction method. Filtered back projection (FBP) meets this requirement, but relies on many high quality X-ray radiographs, which are often not available in a conveyor belt acquisition geometry. On the other hand. iterative reconstruction methods may yield high quality images even with a small number of radiographs, but are orders of magnitude slower. Recently, a neural network FBP (NN-FBP) method was proposed for parallel beam data that proved to be fast and lead to high quality images. (Pelt et al. 2013a) In this work, we present an NN-FBP based CT reconstruction method for inline inspection. Using neural networks, the method computes application specific filters for a Hilbert transform FBP (hFBP) based reconstruction. Results from the proposed neural network based hFBP (NN-hF BP) method on fan beam X-ray radiographs of apples show that, compared to conventional reconstruction methods, NN-hFBP generates images of high quality in a short reconstruction time.