关键字：

摘要：Segmentation of fracture in computed tomography (CT) images of cores is crucial for analyzing rock physical properties. While supervised learning methods perform well in fracture detection, it requires a large amount of labeled data. Unsupervised domain adaptation (UDA) methods are widely used in CT image semantic segmentation tasks, where the key is to efficiently align the feature distributions in the source and target domains. Some unsupervised domain adaptation methods only rely on adversarial learning to achieve feature alignment, making model convergence difficult. Domain shifts exist in CT images of different types of cores due to imaging conditions and geologic features, which can affect the generalization ability of the model. To address these problems, we proposed a new UDA method to segment the fracture, which focuses on overcoming the differences in the noise distribution between domains. It involves two stages: style transfer and multi- source domain adaptation learning. Firstly, our method adopts a style transfer mechanism to align the visual features of the source and target images in order to reduce the differences in gray scale and noise distributions. Secondly, multi-source domain adaptation learning consists of three modules: a modified U-Net segmentation network that extracts robust feature representations by adding attention mechanisms; a domain adaptation module, which utilizes adversarial learning to simultaneously reduce the domain shifts on both high-level semantic features and output space features of images; and a collaborative learning module that corrects the predictions by generating pseudo-labels of the target domain through segmentation networks corresponding to different source domains. Experiments on core datasets from three different geological sources show that our approach gets close to the state-of-the-art in terms of robustness and generalization of fracture segmentation.

卷号：264

期号：

是否译文：否