Research On Table Structure Recognition Algorithms Based On Deep Learning

Table structure recognition is a basic and challenging task in the field of document recognition,which needs to obtain the structure of cells.Recently,deep learning has become the mainstream method for solving the task of table structure recognition,which trains a neural network to get the best fitting model.The algorithm for identifying table rows and columns and the algorithm for identifying table frame lines are the two main algorithms for table structure recognition based on deep learning.The algorithm of identifying table rows and columns identifies the table structure by obtaining the rows and columns of the table.Using semantic segmentation technology to get the rows and columns of the table is an effective scheme.This method simplifies the prediction of rows and columns to the prediction of one column and one row of pixels,which can reduce the difficulty of learning,but causes the problem of low fault tolerance.Table images collected in specific scenes always have complex grid structure,so it is more reasonable to use the algorithm for identifying table frame lines to identify the table structure.Due to the small number of pixels in the table frame lines,this algorithm has the problem of imbalance between positive and negative samples.In order to solve the problems existing in the above algorithms,this thesis makes an intensive study of table structure recognition algorithms,and the specific research contents are as follows:(1)To solve the problem of low fault tolerance caused by simplifying the labels of rows and columns,a table structure recognition network based on the algorithm of identifying table rows and columns is designed in this thesis.The row reconstruction module and column reconstruction module are used to get row features and column features from the features which contain row and column information.In this thesis,the proposed network is tested on ICDAR 2013 dataset.The results show that compared with the algorithm of predicting simplified rows and columns,the average F1 value of rows and columns of the proposed method is higher,reaching 95.24%.(2)In this thesis,the table structure recognition network is redesigned to avoid the row and column reconstruction modules being affected by the detailed information in the low-level features to generate poor attention maps.The row and column reconstruction modules adopt row and column features with more semantic information to guide network learning.And at the decoder,the decoding of row and column features is decoupled.The results on the ICDAR 2013 dataset show that the improved method further improves the performance without increasing the amount of parameters,its row and column average F1 value reaches 95.65%.(3)For the structure recognition of table images collected in the scene,this thesis designs a table structure recognition network based on heatmaps of table frame lines,which is used to solve the problem of positive and negative sample imbalance in the algorithm for identifying table frame lines.The method generates Gaussian heatmaps of the horizontal and vertical frame lines,and employs the encoder-decoder network to regress the probability values of the heatmaps.The results on the bank table dataset with complex and changeable environment show that the method has strong antishadowing,deblurring and anti-tilting ability,and the average accuracy of cells reaches 92.71%.