Table structure recognition based on dual-branch encoder-decoder segmentation

Digitization of paper documents is crucial for the management of modern power grid enterprise. Table structure recognition, which identifies table cells, presents challenges due to diverse table formats. This paper introduces a novel table structure recognition method based on dual-branch encoder-decoder segmentation. The proposed approach converts table structure extraction into row and column segmentation sub-problems, which utilizes a single encoder for feature extraction and two independent decoder branches for segment prediction. In this framework, a Conv-Res-CBAM unit is proposed to enhance feature extraction and transmission. Additionally, the Tesseract OCR engine is incorporated for character recognition. Extensive experiments on two public datasets and a self-collected dataset demonstrate the superiority of our method.