Enhancing seismic image resolution using Brownian diffusion bridge model

Seismic images often lack the high resolution needed for proper identification of subsurface structures and potential hydrocarbon reservoirs, and that is due to factors such as limited data acquisition and the attenuation of seismic waves as they travel through the Earth's subsurface layers. Most of the seismic image enhancement algorithms do not account for prior features of high resolution seismic data. Thus, we propose using a generative diffusion model to enhance seismic image resolution. Specifically, we employ the Brownian diffusion bridge model (BBDM) to translate samples from a low-resolution image distribution to one corresponding to a high-resolution image distribution. To address the issue of training solely on synthetic data and improve the generality of the neural network, we adopted a robust training procedure using the know-distillation technique within a “teacher-student” framework. Field datasets demonstrated the robustness and good performance of the proposed method. Additionally, the intrinsic denoising feature of the diffusion model provides an added image-denoising capability for our methodology.