| Popis: |
Traditional denoising methods often lose details or edges, such as Gaussian filtering. Shearlet transform is a multi-scale geometric analysis tool which has the advantages of multi-resolution and multi-directivity. Compared with wavelet, curvelet, and contourlet transforms, it can retain more edge details while denoising, and the subjective vision and objective evaluation indexes are better than other multi-scale geometric analysis methods. Deep learning has made great progress in the field of denoising, such as U_Net, DnCNN, FFDNet, and generative adversarial network, and the denoising effect is better than BM3D, the traditional optimal method. Therefore, we propose a random noise suppression network ST-hFFDNet based on non-subsampled shearlet transform (NSST) and improved FFDNet. It combines the advantages of non-subsampled shearlet transform, Huber norm, and FFDNet, and has three characteristics. 1) Shearlet transform is an effective feature extraction tool, which can obtain the high and low frequency features of a signal at different scales and in different directions, so that the network can learn signal and noise features of different scales and directions. 2) The noise level map can improve the noise reduction performance of different noise levels. 3) Huber norm can reduce the sensitivity of the network to abnormal data and improve the robustness of network. The network training process is as follows. 1) BSD500 datasets are enhanced by flipping, rotating, scaling, and cutting. 2) AWGN with noise level σ∈[0,75] is added to the enhanced datasets to obtain the training datasets. 3) NSST multi-scale and multi-direction decomposition is performed on each pair of samples of the training datasets to obtain high- and low-frequency images of different scales and directions. 4) Based on the decomposed high and low frequency images, the ST-hFFDNet network is trained by Adam algorithm. 5) All samples of the training data set are carried out in steps (3) and (4), and the trained model is thus obtained. Simulation experiments and real seismic data denoising show that for low noise, the proposed method is slightly better than NSST, DnCNN, and FFDNet and that it is superior to NSST, DnCNN, and FFDNet for high noise. |
