Two-stage domain adaptation for fracture segmentation in electric imaging logging images

Abstract

Electric imaging logging images are widely used for fracture identification due to their high resolution and intuitive characteristics, which is helpful in guiding the development of oil and gas resources. Nevertheless, existing fracture identification methods often face challenges in complex geological backgrounds and when training samples are limited. To address these challenges, this paper presents a two-stage fracture identification method for electric imaging logging images based on domain adaptation. In the first stage, a pseudo electric imaging logging data generator (PEDG) including style transfer algorithm is designed to generate an labeled dataset that is realistic and diverse in fracture morphologies, mitigating the problem of insufficient training samples through image-level domain adaptation. In the second stage, a patch-based output space multi-scale adversarial learning network (pOMAL) is proposed. pOMAL employs adversarial learning between the semantic segmentation model and a multi-scale discriminator (MSD), encouraging the segmentation results of real images resemble those of pseudo-images and improving the generalization ability and noise resistance of the segmentation model through output-level domain adaptation. Experimental results demonstrate that the proposed method achieves better fracture segmentation results in complex electric imaging images with limited training samples.