LGAF‑FracNet: A Dual‑Backbone Deep Learning Framework for Intelligent Fracture Identification in Electrical Imaging Logging

Electrical imaging logging provides rich information on reservoir petrophysical properties and geological features. Fracture identification based on image logs is of great significance for accurate production prediction and reliable estimation of hydrocarbon reserves. However, in electrical imaging log images, fractures typically appear as elongated, low-contrast targets with strong sensitivity to structural continuity. Moreover, variations in formation conditions, imaging parameters, and noise characteristics across different wells pose substantial challenges to existing fracture identification methods, particularly in terms of fine-scale fracture continuity recognition and cross-well generalization. To address these challenges, this study proposes a dual-backbone deep learning framework, termed LGAF-FracNet, for fracture identification in electrical imaging logs. The proposed framework parallelly integrates a convolutional neural network and a Transformer architecture to model local texture features and global semantic relationships, respectively. Considering the circumferential structural characteristics of electrical imaging logs, a liquid ordinary differential equation–based dynamic feature evolution module, an adaptive graph fusion module, and a stripe-aware pooling strategy are incorporated to enhance the representation of elongated and subtle fracture geometries. In addition, a multi-decoder consistency supervision mechanism is introduced to improve cross-well generalization performance. The proposed method is evaluated on a dataset comprising approximately 3,000 electrical imaging log images collected from 21 wells in the Sichuan Basin, covering conductive fractures, resistive fractures, drilling-induced fractures, and bedding structures. A standardized dataset is constructed through manual annotation and data augmentation. Experimental results demonstrate that LGAF-FracNet consistently outperforms mainstream segmentation models in terms of mIoU, F1-score, and pixel accuracy, exhibiting significant advantages in fine-scale fracture continuity, morphological consistency, and cross-well adaptability. These results indicate that the proposed method provides a reliable technical solution for intelligent fracture identification and quantitative characterization in electrical imaging logging.