| Due to the high resolution text scanning system and advanced computer technology,some existing OCR software and recognition systems have been able to help customers with the basic text recognition tasks in most environments.However,only recognition of the text in the table images is not enough.The table structure usually contains importance content,so one of the common practical application scenarios is to recognize the table structure in the document,so as to extract the information in the table.Currently models have made some progress in the table structure recognition task using deep learning methods,but there are still problems in the real world table images.First,there are various types of spatial distortion in real images.In these distortion cases,some assumptions of the model about the table are no longer valid.For example,if the grid of the table is assumed to be horizontal and vertical,the performance of the model will decrease in face of geometric deformation.Second,in the model based on the combination of basic elements,most of the models are combined with text blocks or table units obtained by OCR tools.This combination method regards text blocks as basic units,but the granularity of this basic unit is relatively coarse.A text block may contain the content of multiple table cells.In addtion,when the model models the relationship of each basic element,there is a problem that the relative position of the basic unit is not sufficiently modeled.Aiming at the first problem,this thesis proposes a table structure recognition network with geometry aware ability.On the basis of using the backbone network to extract features,it uses the idea of key point detection to detect vertices and table cell centers.In the geometry aware module,the affine transformation is used to model geometric distortion,so that the model obtains a geometry-aware feature map.On this feature map,the paired regression between the vertices and the center of the table cell is performed,and the structure of the table is restored through post-processing.For the latter two problems,this thesis proposes a table structure recognition network based on position decoupling Transformer.First,the text blocks and its corresponding content are obtained through the OCR tool,and then the text blocks is divided into finergrained text token by using Sentence Piece.Fine-grained text token are combined into a table structure.A Transformer network that decouples content and position is proposed.Each text token is represented by two vectors,a content vector and a position vector,and the two vectors interact through a decoupled attention mechanism.Finally,this thesis uses a supervised contrastive loss to supervise the representation.The method proposed in this thesis proves its effectiveness and rationality through experiments and comparisons. |
