基于低秩近似与自监督学习的地震数据去噪方法研究

Research on Seismic Data Denoising Method Based on Low-Rank Approximation and Self-Supervised Learning

摘要: 地震图像由不同频段的图像组成，提供丰富的地层和纹理信息。然而，地震图像常常受到环境、设备因素干扰，伴随着复杂的随机噪声。传统的地震图像去噪方法依赖约束优化方法，对于先验知识的选择也十分重要。由于受限于手工先验，传统方法的去噪性能急需提升。近来，深度学习被广泛应用于图像处理领域，但是需要训练大量的成对样本，对未经过训练的噪声分布泛化能力差。因此，一种基于优化的方法和深度学习的深度低秩分解的自监督网络的方法——Flex-DLD被提出，通过用一个可训练的神经网络模块(DLD)替代多个复杂的手工求解器，并将其他先验转化为损失函数项，极大地简化了算法的设计和实现流程。本文成功地将基于模型的优化方法与数据驱动的深度学习方法(深度低秩分解)相结合。既充分利用了神经网络强大的非线性表示能力替代传统的低秩分解(如SVD)，又通过WTV来引导网络学习，以更好在去噪的同时保护地震剖面中的结构和边缘信息。

Abstract: Seismic images consist of multi-frequency bands, providing rich stratigraphic and textural information. However, seismic images are often contaminated by environmental and equipment factors, accompanied by complex random noise. Traditional seismic image denoising methods rely on constrained optimization approaches, where the selection of prior knowledge is also crucial. Due to limitations imposed by manual priors, the denoising performance of traditional methods urgently requires improvement. Recently, deep learning has been widely applied in image processing, but it requires training on large amounts of paired samples and exhibits poor generalization ability for untrained noise distributions. Therefore, Flex-DLD—a method combining optimization-based approaches with deep low-rank decomposition via self-supervised networks—has been proposed. By replacing multiple complex manual solvers with a single trainable neural network module (DLD) and converting other priors into loss function terms, the design and implementation of the algorithm are greatly simplified. This paper successfully combines model-based optimization methods with data-driven deep learning methods (deep low-rank decomposition). It fully utilizes the powerful nonlinear representation capabilities of neural networks to replace traditional low-rank decompositions (such as SVD), and uses WTV to guide network learning, so as to better preserve structural and edge information in seismic profiles while denoising.