Research On Frequency Domain High Resolution Radon Transform Inversion Method Based On Neural Network

Research on frequency domain high-resolution Radon transform inversion method based on neural network seismic data processing is a key link in seismic exploration.Radon transform is a common solution method for seismic data processing,and the resolution of radon parameters is a key factor affecting the results of seismic data processing.The traditional Radon transform inversion process involves the inversion of large matrix,which makes the radon inversion process have some problems,such as low computational efficiency and slow convergence speed.For the traditional solution method of Radon transform,based on the iterative soft threshold,this paper introduces the idea of weight matrix in the iterative weighted least square method,so as to introduce the prior information of radon parameters,so as to restrict the inversion process of Radon transform,and overcome the problems of low resolution and slow convergence speed of iterative soft threshold algorithm,A high resolution Radon transform based on iterative weighted soft threshold method is proposed.Radon convolution operator presents mixed phase,while the traditional Radon transform method can not converge to the optimal solution through linear inversion.Therefore,a high-resolution Radon transform inversion method based on convolutional neural network is proposed in this paper.The solution method of high-resolution Radon transform based on one-dimensional convolutional neural network proposed in this paper realizes the mapping from lowresolution radon parameters to high-resolution Radon parameters by virtue of the nonlinear characterization ability of convolutional neural network.In this paper,a series one-dimensional convolutional neural network based on Radon transform deconvolution principle and a parallel one-dimensional convolutional neural network based on residual learning are proposed to realize high-resolution Radon transform from different directions.Then the specific frequency radon parameters obtained by the network are constrained to the inversion of other frequency radon parameters,so as to avoid the problem of low network training efficiency caused by frequency division training.Finally,the high-resolution Radon transform based on convolutional neural network is used for multiple suppression,and compared with the traditional Radon transform solution method,it shows the feasibility and effectiveness of this method.