XYXin-Gang Yang

Xin-Gang Yang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Xin-Gang Yang has a total of 13 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    RTFM-VTB: A Weakly Supervised Video Anomaly Detection Method

    by Cheng Fan, Yuelei Xiao, Zhao Guo, Junyi Li, Peiqi Yang

    With the increasing amount of video data, video anomaly detection has been widely used in security, transportation and other areas. However, weakly supervised video anomaly detection still faces challenges due to problems such as sparse labels and complex and variable samples. To improve the performance of weakly supervised video anomaly detection, we propose a new weakly supervised video anomaly detection method. In this approach, we first extract video features using the Videoswin model instead of the I3D model, which has stronger performance in feature extraction. Next, we introduce the BERT module, which aggregates CNN features from video clips into a single feature, improving classification accuracy and achieving more accurate predictions than methods such as Multiple Instance Learning (MIL). A Transformer Encoder module is also added to model Videoswin features in a more semantically rich way, resulting in a more accurate representation. Our method achieves experimental results of 98.20% on the ShanghaiTech dataset, outperforming existing methods.

  • Open Access

    Express-related Counterfeit Cigarette Crime Prediction with Imbalanced Data-based Machine Learning Techniques

    by Wei Liu, Rong-Yao Xu, Lang-Chao Qiao, Jin-Lu Wang, Bao-Hong Gao, Xin-Gang Yang, Wen-Tao Feng

    The number of cigarette-related packages and normal packages is often extremely unbalanced in practice, this characteristic must be considered in the express-related counterfeit cigarette crime prediction, otherwise the performance of prediction model on unknown packages will be affected badly. In this paper, an approach of using SMOTE-ENN technology to deal with imbalanced express data and establish the express-related counterfeit cigarette crime prediction model is proposed. In the preprocessing stage, original data sets are structured by Chinese word segmentation method. In feature calculation and analysis stage, the spatio-temporal features of express data are calculated and analyzed, and the optimal features are further obtained by feature reduction algorithm named PCA. Next, based on optimal features, balanced data are obtained by mixed sampling. Finally, the classifier model was trained and optimized to predict the probability of a package containing counterfeit cigarettes in express data including 2851 normal samples and 285 cigarette-related samples randomly selected from 3077 cigarette-related samples. The performance of all classifiers is improved after using the SMOTE_ENN to balance datasets, and the indicators are above 98%. Experimental results demonstrate feasibility, effectiveness of the proposed approach.


  • Open Access

    Handwritten Text Recognition of Imagery Acquisition Sheet in Tobacco Industry Using an Improved CRNN Network

    by Wei Liu, Yan Wang, Lang-Chao Qiao, Jin-Lu Wang, Yi-Ran Wei, Hai Liu, Xin-Gang Yang, Wen-Tao Feng

    In the tobacco industry, the acquisition sheet is a kind of document used by the tobacco monopoly administration to record the information about the acquisition of fake and illicit cigarettes when the tobacco crime is seized by the "Internet Plus Express". To automatically identify the information helps to realize the automation of information management of tobacco monopoly administration department, and then build delivering for smoke crime management and prediction system, thus to deliver smoke involved crime effectively, and realize the wisdom of the tobacco industry monopoly, so to handwriting recognition research of the acquisition sheet image is very necessary. In this paper, an improved CRNN network text recognition method is proposed. Firstly, the acquisition sheet image with interference such as shadow and seal is preprocessed based on color channel separation method. Then, in the presence of printed text, the YoloX model is used to detect the handwritten text in the image, and the handwritten text line dataset is obtained. Then, the CNN network is improved by adding Swish activation function and attention mechanism to obtain the depth image features of handwritten text line data. BiLSTM network predicts the input feature sequence and finally realizes the recognition of handwritten text. Through the algorithm in this paper, the detection of handwritten text in the acquisition sheet image reaches an average accuracy of 0.983. The experimental results show that, compared with the CRNN model, the improved CRNN model in this paper has a better effect on the recognition of handwritten text in the acquisition sheet image.