基于残差Haar离散小波变换与多约束损失的深度学习地震超分和去噪方法研究OA

Addressing the pervasive issues of insufficient resolution and noise interference in field seismic data,this study develops a deep learning approach that integrates architectural enhancements with multi-constrained loss function optimization to achieve joint seismic image super-resolution and denoising.Built upon the U-Net framework,the method replaces standard downsampling operations with a Residual Haar Discrete Wavelet Transform(RHDWT)module to better preserve the frequency characteristics of seismic signals.Simultaneously,sub-pixel convolutional layers are employed for efficient upsampling,complemented by residual blocks to enhance the network's ability to learn detailed seismic features.Regarding the loss function design,a composite function comprising L1 loss,Multi-Scale Structural Similarity(MS-SSIM)loss,and Frequency Focal Loss(FFL)is constructed.This framework constrains the training process across multiple dimensions—image reconstruction error,structural consistency,and spectral information—thereby synergistically optimizing both super-resolution and denoising performance.Experimental results on both synthetic and field data demonstrate that the proposed method effectively suppresses noise while enhancing the resolution of seismic profiles.It successfully recovers high-frequency details,improving the lateral continuity of weak reflectors and the spatial identification of thin-bed structures.