Fracture recognition in ultrasonic logging images via unsupervised segmentation network

Image well logging is an intuitive approach to identify fractures of reservoir for oil and gas exploration. However, these logging images are rare and nonannotated. A method of unsupervised segmentation network based on convolutional neural network is adopted to automatically extract pixels pertaining to fracture information in this paper. We propose a modified model to accomplish domain adaptation from the source domain with similar fractures information to the target domain, which can improve the accuracy of fracture recognition. The network is trained in the source domain with ground truth and tested in the target domain without any labels. Compared with the experimental results of other classical methods, this method has demonstrated satisfactory performances in terms of accuracy and visual quality even if the logging image dataset is insufficient.

Automated summary

The paper introduces an unsupervised segmentation network based on convolutional neural network to automatically extract pixels pertaining to fracture information in logging images. It proposes a modified model for domain adaptation from source domain to target domain to improve fracture recognition accuracy. The network is trained in the source domain with ground truth and tested in the target domain without labels, showing satisfactory performance compared to other classical methods.