Research on Fracture Recognition in Well Logging Images: Adversarial Learning with Attention

Semantic recognition of fractures in well logging images is vital for engineers to implement oil and gas exploration. An essential approach to accomplishing the object is to adopt deep learning based on convolutional neural networks. However, due to the lack of annotated labels in well logging images, it is scarcely available to directly train the semantic segmentation network. In this paper, we explore a domain shift model attempting to achieve domain adaptation from one annotated dataset to our target well logging images. This model's core is to utilize adversarial learning, including generator and discriminator, which can prompt the model to generate fracture segmentation similar to the source domain in the target domain. For enhancing the domain adaptive model further, an attention module is introduced, which can suppress redundant noise in semantic segmentation results. We demonstrate that our proposed model performs well by extensive tests and ablation experiments in the ultrasonic well logging images.