LRATNet: Local-Relationship-Aware Transformer Network for Table Structure Recognition

Guangjie Yang, Dajian Zhong, Yu-Jie Xiong, Hongjian Zhan

Abstract: Table structure recognition is a challenging task due to complex background and various styles of tables. Existing methods address this challenge by exploring adjacency relationship prediction, image-to-text generation, logical position prediction, etc. However, these methods either adopt Graph Convolutional Network (GCN) structures, which mainly focus on the local context information, or Multi-Head Attention (MHA) structures, which mainly focus on the global context information. Both of them ignore the correlation between local and global features. In this paper, we propose a Local-Relationship-Aware Transformer Network (LRATNet) for table structure recognition. LRATNet constructs a robust correlation between local and global information using the LRAT module. The LRAT model has been adapted into three distinct variants: Row-LRAT, Col-LRAT, and Spa-LRAT. These variants are designed to emphasize specific aspects of information: row information, column information, and spatial information, respectively. This is achieved through the exploration of different adjacency relationships. This improves the performance of logical location prediction. Additionally, we have developed a new loss function called Lstage, which is designed to improve accuracy in predicting logical positions. Experimental results demonstrate that our method outperforms existing approaches on three public datasets.