Abstract:In the new type of power system, the application of large models is gradually increasing, and they have played a huge role in processing and analyzing a large amount of power data, and improving the efficiency of power grid operations. To solve the normalization problem of tables in the model training data, this paper proposes a deep learning method that combines Spatial CNN and Grid CNN for the identification and reconstruction of table structures. The method predicts the split line mask through the Spatial CNN and extracts the split lines by using the connected component analysis algorithm (CCA) to build a grid of cells. The Grid CNN cell merging module merges adjacent cells to correct splitting errors and form an accurate table structure. To improve the efficiency of the model, a feature pyramid network (FPN) based on ResNet-18 is used as the backbone network to integrate multi-scale feature information and enhance the recognition ability of the table structure. This method not only improves the accuracy of table recognition but also optimizes the reconstruction process, providing an effective technical solution for automated table processing. The experimental results show that compared with the existing excellent performance methods, the table structure recognition method in this paper has increased by 1.1 percentage points and 0.3 percentage points in precision and F1 metrics on the SciTSR dataset; on the more complex table dataset PubTabNet, the TEDS-Struct metric of the table structure recognition method in this paper reached 97.1%. The research in this paper provides a high-efficient and accurate solution for table processing and training of large language models.