The quality of seismic images is often affected by the limitation of acquisition conditions and the interference of noises, which causes the low resolution of seismic images and misleads the following geological interpretation. Although the super-resolution method for seismic images based on convolutional neural network (CNN) has behaved well, the quality of weak events especially deep events is still need to be improved, due to CNN is limited by the receptive fields, which results in weaker ability to perceive relationships among pixels far apart. In this letter, we solve this problem by designing a combination network of CNN and transformer (CNCT). CNCT consists of three parts, edge feature fusion block (EFB), deep feature mining block (DMB), and feature enhancement block (FEB). The EFB aims to fuse the input low-resolution (LR) image and the corresponding edges obtained by the Sobel algorithm and performs preliminary shallow feature extraction. DMB mines deeper features by stacking residual blocks, and each residual block makes full use of its excellent perception of global and local information by combining transformer and CNN. Finally, the FEB uses subpixel convolution for upsampling to expand the size of feature maps. The experimental results on synthetic data and field data show that CNCT not only behaves better on perception effect and texture details than that of other deep learning (DL) methods but also can suppress noise and improve the dominant frequency.

Full waveform inversion (FWI) is currently the most accurate technique for obtaining the properties of subsurface media. The absence of low frequencies in the observed data caused cycle-skipping phenomenon and poor initial model which affect the convergence of FWI. We propose a global-feature-fusion and multi-scale network (GM-Net) in a way of supervised learning to compensate for the absent low frequency components in the observed data trace by trace. The difficulty of extrapolating frequency is to achieve smoothness and continuity when changing from high frequency signals to low frequency signals, which is visually shown in the reduction and movement of the sidelobes in high-frequency signals and the overall oscillation of the signals is slowed down. For achieving better extrapolation, the encoder-decoder architecture with multi-scale feature extraction is designed as the backbone of the network. For avoiding the loss of information, we propose to perform 1/2 down-sampling on the original input signal separately based on the odd and even time samples, and then concatenate them along the channel dimension. Since 1-dimensional (1D) seismic data is a type of time-series signal and the wavelengths of low frequencies are long, we pay more attention to the relevance of contextual information. Thus, dilated convolution layers, gridding convolution blocks and non-local attention blocks are used to enlarger the receptive field both in time and channel dimensions to extract and fuse global features. Numerical tests both on synthetic data and different types of field marine data demonstrate the feasibility and generalization of our method.

Recorded seismic data often is characterized by a low signal-to-noise ratio (SNR) that can hinder subsequent imaging and interpretation tasks. Thus, it is necessary to explore a method to recover weak signals from strong background noise. While numerous studies have demonstrated the effectiveness of deep-learning methods in seismic noise attenuation, enhancing their capability to recover weak signals under low SNR conditions remains an area for further exploration. To address this issue, we propose an edge-feature-guided wavelet U-Net (EFGW-UNet). In this novel architecture, we utilize the discrete wavelet transform to replace the pooling operation deployed in the conventional U-Net, thereby maintaining more detailed information of effective signals. Meanwhile, we also design a dual decoder for edge detection to obtain the shape and edge information on the effective signals. Finally, to fuse multi-level image features and edge features, an attention feature fusion module is deployed. In the experimental part, we use synthetic and real data to illustrate the effectiveness of EFGW-UNet. Our results suggest better denoising performance than competitive methods, especially for weak signal recovery submerged in heavy noise.

In seismic exploration, erratic noise is a type of intense and complicated interference with large-amplitude and non-Gaussian distributions. The presence of erratic noise has been demonstrated to corrupt reflection events, adversely affecting the identification of effective signals. Nonetheless, conventional and time-frequency thresholding denoising methods based on the least-squares scheme usually assume that the seismic random noise has a Gaussian distribution, which is not the case for erratic noise. Therefore, the attenuation for erratic noise is challenging, owing to the deviation from the assumptions of the conventional methods. The recent application of convolutional neural networks (CNNs) to seismic data processing has yielded promising results. However, these CNN-based frameworks always have limited feature-interaction capability, resulting in the degeneration in denoising performance when coping with intense erratic noise. To address this issue, a novel joint-guided denoising network (JGD-Net) is proposed in this study. Unlike conventional CNN frameworks, JGD-Net uses a joint-guided scheme and attention mechanism to enhance the denoising capability. We generate synthetic records using published geological models such as Marmousi and salt dome to compose our training dataset. Furthermore, a novel loss function based on L1 norm and hyperbolic tangent function is designed to further ensure the optimization process of the training procedure and ease the influence of abnormal energy of erratic noise. Both synthetic and field data are processed sfor the evaluation of denoising performance. Compared with other popular methods, JDG-Net shows advantages in attenuating intense erratic noise, particularly under extremely low signal-to-noise ratio (SNR) conditions.

The super-resolution reconstruction for seismic images obtained by multistep processing of field data is essential due to the noise contamination, sparse geometry, and low dominant frequency of events, which impairs the subsequent seismic interpretation. Deep learning-based methods show strong potential in super-resolution through supervised learning. Generative adversarial networks (GANs) have shown capability in super-resolution of different kinds of images; however, it is limited in enhancing the detailed geological structures of seismic images that are fatal for interpretation. To address this issue, we propose a super-resolution hybrid bilateral attention GAN (SHBGAN) to improve the recovery of weak signals and the reconstruction of geological structures. Specifically, the generator employs hybrid and bilateral attention modules (BAMs) to enhance the capture ability of global and local features. Meanwhile, we use dilated convolutional layers instead of batch normalization (BN) layers in the residual block to improve the generalization ability of the trained model. Meanwhile, the discriminator employs global average pooling and convolutional layers to score the authenticity of seismic images rather than the probability to enhance the stability of training. In addition, we add the mean structural similarity (MSSIM) term to the loss function of generator to improve the perception quality of predictions. The numerical tests on both synthetic and field data show that SHBGAN is more effective than competing methods in recovering weak signals and reconstructing subtle faults.

Seismic data reconstruction is always an essential step in the field of seismic data processing. Effective reconstruction methods can obtain high-density information at low-cost and also recover missing seismic data. Due to the strong feature extraction ability, convolutional neural network (CNN) has shown remarkable performance in numerous fields of data processing and been gradually applied to seismic data reconstruction. However, most of CNN-based methods applied to seismic data reconstruction only consider features in single scale or just utilize simple interactions between different scales, which is likely to result in performance degradation when facing complex and extremely incomplete seismic data. To further promote the performance of CNN-based methods in seismic data reconstruction, a novel multiscale enhanced attention network (MSEA-Net) is proposed based on the self-enhanced scheme. In general, MSEA-Net has a multiscale architecture which can significantly improve the processing accuracy by fusing the potential features in different-resolution seismic data. From the basis, a parallel sparse residual block is designed and applied in MSEA-Net to enhance processing efficiency and avoid overfitting issues. In addition, a dense spatial attention block is also introduced to the network to further reinforce the effective features, thereby strengthening the reconstruction performance. Experimental results demonstrate that our proposed network can effectively reconstruct incomplete seismic data including regular missing data, irregular missing data, and even consecutively missing data with big gap, which is superior than exist interpolation methods including commonly used U-Net.

Distributed acoustic sensing (DAS) has been considered as a breakthrough technique in seismic data collection owing to its advantages in acquisition cost and accuracy. However, the existence of complex background noise combined with a tough exploration environment always results in incomplete data with a low signal-to-noise ratio, posing a big challenge for the subsequent processing of DAS data. To improve the quality of DAS data, convolutional neural networks (CNN) have gradually been utilized to deal with the denoising and reconstruction tasks. Meanwhile, some successful applications have verified that CNN-based methods can significantly alleviate the impacts of DAS background noise and missing trace records, compared with conventional approaches. Nonetheless, in most researches, the denoising and reconstruction tasks are accomplished independently, severely affecting the processing efficiency. In this study, a multi-scale encoder-decoder network (MEDN) is proposed to simultaneously achieve the DAS background noise suppression and weak signal recovery through a unified model. Generally, MEDN can extract the different-scale features through both a multi-scale network architecture and a multi-scale residual (MSR) block. The captured different-scale features are then fused to enhance the effective feature. In addition, the encoder-decoder scheme is also utilized in the design of the network architecture to further enhance the reconstruction performance. Moreover, depthwise separable convolution (DSC) blocks are also utilized to ease the computational burden and improve the processing efficiency. Theoretical and field data processing results show that MEDN can provide better denoising and reconstruction performance than conventional methods and popular CNN-based frameworks.

Nowadays, there are plenty of works introducing convolutional neural networks (CNNs) to the steganalysis and exceeding conventional steganalysis algorithms. These works have shown the improving potential of deep learning in information hiding domain. There are also several works based on deep learning to do image steganography, but these works still have problems in capacity, invisibility and security. In this paper, we propose a novel CNN architecture named as ISGAN to conceal a secret gray image into a color cover image on the sender side and exactly extract the secret image out on the receiver side. There are three contributions in our work: (i) we improve the invisibility by hiding the secret image only in the Y channel of the cover image; (ii) We introduce the generative adversarial networks to strengthen the security by minimizing the divergence between the empirical probability distributions of stego images and natural images. (iii) In order to associate with the human visual system better, we construct a mixed loss function which is more appropriate for steganography to generate more realistic stego images and reveal out more better secret images. Experiment results show that ISGAN can achieve start-of-art performances on LFW, PASCAL-VOC12 and ImageNet datasets.