Reconstruction of Borehole Image Gaps via Adversarial Edge Learning

Borehole image gaps severely disrupt the continuity of formation textures, limiting the accuracy of reservoir parameter inversion and compromising the fidelity of geological interpretation. Traditional interpolation techniques, such as kriging and morphological inpainting, often produce blurred edges and introduce artifacts, whereas existing deep learning-based methods that rely on randomly generated training masks frequently fail to align with the actual distribution of image gaps. To overcome these limitations, a high-precision borehole image reconstruction method based on adversarial edge learning is proposed, enabling the accurate restoration of geological structures. The proposed approach includes three core components. The first is a dynamic mask generation strategy that uses non-random lateral translation and longitudinal overlapping cropping to construct geologically representative datasets. The second is a two-stage adversarial EdgeConnect framework constrained by multiple loss functions, including L1 loss, perceptual loss, style loss, and total variation loss, to maintain both local texture fidelity and global structural consistency. The third involves training the reconstruction network using the tailored dataset and deep learning model. Experimental evaluations show that our method outperforms Deep Generative Prior (DGP), Deep Image Prior (DIP), and Generative Multi-column Convolutional Neural Network (GMCNN) in terms of Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR), with reduced local reconstruction errors and improved variance explanation. The reconstructed images better preserve fracture morphology and texture continuity, enhancing their geological plausibility. Field applications confirm that the proposed method provides a high-fidelity data foundation for seismic inversion and reservoir modeling, offering substantial value for engineering and interpretation tasks.