Ring artifacts segmentation on microtomographic images by convolutional neural networks

Ring artifacts in X-ray microtomographic images can lead to errors in the construction of digital twins of rock samples for flow simulation. Previously, we considered an algorithm for detecting ring artifacts by means of matching filtering of image slices in a polar coordinate system. However, that approach is inapplicable for an arbitrary fragment of an image and requires adjustment of parameters from image to image. In this paper, we propose the segmentation method based on convolutional neural network. Two network architectures are considered: SegNet and U-net. To create a big and representative training and validation datasets, we propose an algorithm for transferring ring artifacts detected by the existing approach from one image to another. Our task-specific data augmentation improves outcomes in comparison with conventional augmentation techniques. The trained model successfully segments ring artifacts even for sample images and artifacts that were not in the training set. The developed algorithm is used to assess the quality of microtomographic images and local correction of image regions damaged by ring artifacts.