基于两阶段深度学习的表格结构识别方法

Table Structure Recognition Method Basedon Two-stage Deep Learning

鉴于在图像中识别表格结构面临着表格样式众多、图像质量各异等难题,提出一种融合表格线与文字块信息的两阶段深度学习框架,以实现少线复杂表格结构的识别.首先,将残差结构引入U-Net语义分割网络中,增强网络传递表格线信息能力,完成表格线的识别;然后,加入文字块位置信息以提高模型识别无线或少线表格结构的能力.该方法在PubTabNet数据集上的树编辑距离(tree-edit-distance similarity,TEDS)评分达到95.95.实验证明,该方法在识别少线表或无线表时表现优秀,并能高效、准确地识别存在合并单元格的复杂结构表格.

In view of the difficulties of table structures recognition in images,such as numerous table styles and different image quality,a two-stage deep learning framework integrating table lines and text block information was proposed to realize the recognition of complex table structures with few lines.Firstly,the residual structure was introduced into the U-Net semantic segmentation network to enhance the network transmission ability of table line information and complete the recognition of table line.Then,the text block location information was added to improve the ability of the model to recognize wireless or less linear table structures.The TEDS(tree-edit-distance similarity)score of this method was 95.95 on PubTabNet dataset.The experimental results suggested that the proposed method performed well in recognition of few line tables or wireless tables,and could efficiently and accurately recognize the complex structure tables with merged cells.