PGSNet : Improve table cell row-column structure detection in YOLOv8

For the intelligent recognition of complex table structures facing dense cells, local details are difficult to extract, and the computational complexity of long-distance table frame line extraction is too high. In this paper, we propose a table data detection algorithm, PGSNet, based on the YOLOv8 framework. Firstly, the C2fPSP module is designed in the C2f structure, and the PSP module is constructed from the windmill shape convolution PConv and SPConv instead of the original Bottleneck module in order to mine the detailed features of the dense table in a lightweight way; To address YOLOv8’s limitations in table structure recognition—including insufficient local-global feature integration, constrained multi-scale detection, and channel redundancy—this paper proposes the Global Dilated Feature Aggregation module. By synergizing dilated convolutions with a self-attention mechanism to achieve dynamic multi-scale perception, combined with channel shuffling to enhance feature diversity; moreover, the sparse self-attention mechanism is introduced into the backbone network, which dynamically divides the feature region through the multi-scale sparse rate. A sparse self-attention mechanism is introduced to dynamically divide the feature regions through a multi-scale sparse rate to reduce the redundant computation of irrelevant regions, and at the same time, the sensitivity of the model to local artifacts in the corners of the form is strengthened to enhance the robustness to the frame line structure of complex forms. Experimental results on the ICDAR19-cTDaR dataset show that this paper’s method achieves the SOTA of the cell detection task with 93.1%, 83.1%, and 87.8% for the Precision, Recall, and F1 score, respectively, for the SOTA of this dataset; and the F1 score of the row-column detection line of the TabStructDB dataset reaches 93.3%, which is 3.4% higher than that of the baseline model YOLOv8 by 3.4%, which also has a significant advantage on this dataset, verifying its efficiency and accuracy in dense table structure recognition.