In this paper, we propose a simple yet effective self-supervised method called spatio-temporal contrastive learning (ST-CL for 3D skeleton-based action recognition. ST-CL acquires action-specific features by regarding the spatio-temporal continuity of motion tendency as the supervisory signal. To yield effective representations, ST-CL first designs a novel contrastive proxy task by providing different spatio-temporal observation scenes for the same 3D action and pulling them together in the embedding space. Second, three key components are devised in the action encoding to efficiently extract representations in contrastive tasks: (1 Information Representation introduces the awareness of joint type when analyzing the motion dynamics. (2 Non-local GCN learns a data-driven graph topology structure and promotes a spatial message passing among long-range joints in each frame. (3 Multi-Scale TCN makes larger receptive fields for capturing richer longe-range temporal dynamics amomg adjacent frames. In ST-CL, the effective proxy tasks yield useful representations and efficient action encoding further enhances representation capacity. \revise{As validated on the four large-scale datasets, ST-CL is a strong baseline with high performance and efficiency for the contrastive learning study of the skeleton data. Compared to previous self-supervised methods, the proposed ST-CL achieves significant improvement consistently with a smaller model size and better training efficiency.

Time-frequency analysis (TFA) is widely used to describe local time-frequency (TF) features of seismic data. Among the commonly used TFA tools, sparse TFA (STFA) is an excellent one, which can obtain a TF spectrum with good readability. However, many STFA algorithms suffer from expensive calculation time and unavoidable prior knowledge, such as the iterative shrinkage-thresholding algorithm (ISTA) and the sparse reconstruction by separable approximation (SpaRSA). Inspired by the unrolled algorithm and its successful applications in signal processing, we propose a deep learning-based ISTA unrolled algorithm, which is named the sparse time-frequency analysis network (STFANet). The STFANet contains two parts, i.e., the sparse time-frequency spectrum generator and the reconstruction module. The former learns how to transform a one-dimensional (1D) seismic signal from a large amount of unlabelled data into a two-dimensional (2D) sparse time-frequency spectrum, which is implemented based on the proposed unrolled iterative dynamic shrinkage-thresholding (UIDST) algorithm. Note that the UIDST algorithm is carried out by using a simplified deep learning network. The latter serves as a physical constraint of model training to ensure that our generator obtains an accurate TF spectrum, which is actually an inverse time-frequency transform. In this study, the traditional inverse short-time Fourier transform (STFT) is utilized in the reconstruction module. To test the effectiveness of the proposed model, we apply it to 3D post-stack field data. The results show that, compared with the traditional TFA tools, the STFANet can availably compute time-frequency spectrum with better readability, which benefits seismic attenuation delineation.

Seismic facies classification plays an important role in oil and gas reservoir interpretation. In the past few years, convolution neural network (CNN)-based models have been widely used in supervised seismic facies classification. However, to improve some inherent limitations of CNNs, it is significant to explore alternative architectures, such as the Transformer with the self-attention mechanism. In this study, based on the Segformer, a cutting-edge semantic segmentation Transformer, we propose a U-shaped model of joint the Segformer and the Hypercolumn representation for seismic facies classification, named U-Segformer-Hyper. As the emerging lightweight variant of the Transformer for seismic facies segmentation, the proposed U-Segformer-Hyper consists of a specific encoder-decoder-hypercolumns scheme, which can extract various features in different layers and fuse the output features of different layers with different scales. In the application of the public F3 seismic data, compared with the CNN Benchmark model, the U-Segformer-Hyper model has fewer parameters, fewer floating-point operations per second, and higher accuracy of classification in both the section-based and patch-based training modes. Moreover, the proposed U-Segformer-Hyper is open-source, which benefits to further explore alternative deep models in seismic interpretation.

Synchrosqueezing transform (SST) is an effective time-frequency analysis (TFA) approach for the processing of nonstationary signals. The SST shows a satisfactory ability of the TF localization of the nonlinear signal with a slowly time-varying instantaneous frequency (IF). However, for the signal of which ridge curves in the TF domain are fast varying, or even almost parallel to the frequency axis, the SST will provide a blurred TF representation (TFR). To solve this issue, the transient-extracting transform (TET) was recently put forward. The TET can effectively characterize and extract transient features in the much concentrated TFR for the strongly frequency-modulated (FM) signal, especially the impulse-like signal. However, contrary to the SST, it is not suitable for weak FM modes. In this study, we propose a TFA method called the time-synchroextracting general chirplet transform (TEGCT). The TEGCT can achieve a highly concentrated TFR for strong FM signals as well as weak FM ones. Quantized indicators, the concentration measurement and the peak signal-to-noise ratio, are used to analyze the performances of the proposed method compared with those of other methods. The comparisons show that the TEGCT can provide a result with better TF localization. Then, the proposed method was applied to the spectrum analysis of the seismic data for oil reservoir characteristics. The horizontal slices of the offshore 3-D seismic data show that the TEGCT delineates more distinct and continuous subsurface channels in a fluvial-delta deposition system. All the results illustrate that our proposed method is a good potential tool for seismic processing and interpretation in the geoscience.

Robust point set registration is a challenging problem, especially in the cases of noise, outliers, and partial overlapping. Previous methods generally formulate their objective functions based on the mean-square error (MSE) loss and, hence, are only able to register point sets under predefined constraints (e.g., with Gaussian noise). This article proposes a novel objective function based on a bidirectional kernel mean p-power error (KMPE) loss, to jointly deal with the above nonideal situations. KMPE is a nonsecond-order similarity measure in kernel space and shows a strong robustness against various noise and outliers. Moreover, a bidirectional measure is applied to judge the registration, which can avoid the ill-posed problem when a lot of points converges to the same point. In particular, we develop two effective optimization methods to deal with the point set registrations with the similarity and the affine transformations, respectively. The experimental results demonstrate the effectiveness of our methods.

3D human action recognition has received increasing attention due to its potential application in video surveillance equipment. To guarantee satisfactory performance, previous studies are mainly based on supervised methods, which have to add a large amount of manual annotation costs. In addition, general deep networks for video sequences suffer from heavy computational costs, thus cannot satisfy the basic requirement of embedded systems. In this paper, a novel Motion Guided Attention Learning (MG-AL) framework is proposed, which formulates the action representation learning as a self-supervised motion attention prediction problem. Specifically, MG-AL is a lightweight network. A set of simple motion priors (e.g., intra-joint variance, inter-frame deviation, intra-joint variance, and cross-joint covariance), which minimizes additional parameters and computational overhead, is regarded as a supervisory signal to guide the attention generation. The encoder is trained via predicting multiple self-attention tasks to capture action-specific feature representations. Extensive evaluations are performed on three challenging benchmark datasets (NTU-RGB+D 60, NTU-RGB+D 120 and NW-UCLA). The proposed method achieves superior performance compared to state-of-the-art methods, while having a very low computational cost.