Latent Progressive Diffusion Model for Super-Resolution Reconstruction of Oil and Gas Well Perforation Images

Abstract

To address the challenges in super-resolution (SR) reconstruction of oil and gas well perforation images—such as excessive GPU memory consumption during high-resolution training, susceptibility to losing complex textural details, severe artifact generation, low computational efficiency of the original Denoising Diffusion Probabilistic Model (DDPM) in pixel space, and the failure of naive progressive diffusion to resolve computational costs from a dimensional perspective—this paper proposes a latent progressive denoising diffusion probabilistic model for super-resolution (LP-DDPMSR). The model integrates the dimensional compression advantages of the Latent Diffusion Model (LDM) with the staged training strategy of progressive diffusion. By mapping perforation images into a low-dimensional latent space via a pre-trained Variational Autoencoder (VAE) for the diffusion process, the approach fundamentally alleviates the GPU memory burden caused by pixel-level computations. A Latent Perforation Hierarchical Feature Enhancement Sub-Network (L-PFEN) is designed to specifically optimize the latent representation of high-frequency and low-frequency features in perforation images. The Latent Perforation-Adapted U-Net (L-PA-U-Net) is enhanced by embedding a Time-Aware adaptive Cross-Attention (TACA) mechanism, which strengthens feature modeling of critical perforation regions within the latent space. Furthermore, a three-stage progressive training framework in the latent space is established to achieve a smooth resolution escalation from 64×64 to 256×256. Experimental results demonstrate that, under 2×, 4×, and 8× magnification factors for oil and gas well perforation images, LP-DDPMSR outperforms Bicubic, ESRGAN, DDPMSR, and LDM-based methods in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). At 8× magnification, the model achieves a PSNR of 20.83 dB, an SSIM of 0.8217, and a Fréchet Inception Distance (FID) of 45.29. Compared to the pixel-space progressive model DDPMSR, LP-DDPMSR reduces training GPU memory consumption by 68.5% and improves training efficiency by 62.3%. While ensuring high-fidelity restoration of perforation image details, the model significantly enhances computational efficiency, providing more effective technical support for oil and gas well perforation quality assessment.