Text Recognition of Measurement Experiment Detection Table Based on Feature Extraction and Feature Fusion

The text data for quantitative testing contains a large number of tables, and the main information is stored in the tables. This article proposes an improved feature extraction and fusion model for identifying the complex and diverse table structures, as well as a large number of merged and blank cells, in the text of power metering test detection. By using CVT for feature extraction, we can better detect blank cells and merged cells, improve model accuracy, and use fully connected layers instead of semantic segmentation layers for row and column detection. At the same time, we remove the complex post-processing steps based on semantic segmentation methods, which improves model speed. In order to improve the effect of cell merging, the proposed model uses quadratic ROI+CVT+SCNN to enhance the information exchange ability between adj acent cells. The final experimental results indicate that the proposed model has certain advantages in speed and accuracy compared to existing models, and can solve problems such as blank cells.