Unsupervised Deep Learning for Laboratory-Based Diffraction Contrast Tomography

An important leap forward for the 3D community is the possibility to perform non-destructive 3D microstructural imaging in the home laboratories. This possibility is profiled by a recently developed technique—laboratory X-ray diffraction contrast tomography (LabDCT). As diffraction spots in LabDCT images are the basis for 3D reconstruction of microstructures, it is critical to get their identification as precise as possible. In the present work we use a deep learning (DL) routine to optimize the identification of the spots. It is shown that by adding an artificial simple constant background noise to a series of forward simulated LabDCT diffraction images, DL can be trained and then learn to remove high frequency noise and low frequency radial gradients in brightness in the real experimental LabDCT images. The training of the DL routine is unsupervised in the sense that no human intervention is needed for labelling the data. The reduction in high frequency noise and low frequency radial gradients in brightness is demonstrated by comparing line profile scans through the experimental and the DL output images. Finally, the implications of this reduction procedure on the spot identification are analysed and possible improvements are discussed.