Seismic Data Reconstruction Based on a U-Net Network with Fused Perceptual Loss

In seismic exploration, due to the limitations of terrain conditions and environmental factors, the acquired seismic data often have irregular distribution and missing phenomena, which will adversely affect the subsequent data processing and interpretation. Therefore, seismic data reconstruction becomes an important step to improve data quality and analysis accuracy. In order to improve the detail recovery ability of seismic data interpolation reconstruction, a U-Net network fused with perceptual loss is constructed for the recovery and detail reconstruction of undersampled seismic data. This method takes U-Net as the basic framework, and uses its encoder-decoder structure to capture multi-scale spatial information in seismic data. In the training process, the VGG Loss is introduced and the mean square error is combined to construct a hybrid loss function to optimize the network’s ability to perceive the details of the image structure. The training sample consists of standard seismic image slices, which are divided into training sub-blocks by a sliding window. The experimental results show that the U-Net model with perceptual loss is superior to the traditional U-Net in the quantitative indicators of Mean Square Error (MSE), Signal-to-Noise Ratio (SNR) and Structural Similarity index (SSIM). At the same time, it performs more clearly in the recovery of structural details such as faults and interfaces, which effectively alleviates the problem of excessive smoothing. The introduction of perceptual loss improves the expression ability of the model for the spatial structure of seismic data, and provides a more advantageous solution for the reconstruction of undersampled seismic data.