Seismic Random Noise Attenuation Using Non‐Local Reference‐Guided Deep Image Prior

ABSTRACT Supervised deep learning methods have been widely applied for seismic random noise attenuation, but their dependence on large volumes of clean training data limits their practicality. Deep image prior (DIP) provides an unsupervised alternative by exploiting the structural bias of convolutional neural networks. However, its performance is sensitive to the choice of stopping iteration and does not explicitly incorporate structural prior information inherent in seismic data. In this study, we propose a non‐local reference‐guided deep image prior framework for seismic random noise attenuation. Non‐local self‐similarity (NSS) is extended from the patch level to the pixel level to improve noise level estimation accuracy and to generate structurally consistent reference data. Based on the estimated global noise level, a noise‐driven early stopping criterion is introduced to determine the termination point of DIP optimization in a fully unsupervised manner. The NSS‐refined reference is used as the network input, allowing structural information to be incorporated into the reconstruction process. In addition, selective weight decay applied to the decoder layers further enhances the separation between signal and high‐frequency noise. Experiments on synthetic and field seismic data indicate that the proposed method effectively attenuates random noise while preserving structural continuity and reflection characteristics. Compared with existing unsupervised approaches, the method provides more stable optimization behaviour and improved reconstruction quality without requiring clean training data.