Table structure understanding and its performance evaluation

With the large number of existing documents and the increasing speed in the production of new documents, finding efficient methods to process these documents for their content retrieval and storage becomes critical. Tables are a popular and efficient document element type. Therefore, table structure understanding is an important problem in the document layout analysis field. This paper presents a table structure understanding algorithm using optimization methods. It includes steps of column style labeling, large horizontal blank block equivalence subsets location, statistical refinement, iterative updating optimization and table decomposition. The column style labeling, statistical refinement and iterative updating optimization steps are probability based, where the probabilities are estimated from geometric measurements made on the various entities with which the algorithm works in a large training set. Each step of our table structure understanding algorithm has some tuning parameters. We initially set the parameters with some conjectural values. Then with a global parameter optimization scheme, we update these values using a line search optimization algorithm. We use a performance evaluation protocol employing an area overlapping measure. With this scheme, we can obtain statistically satisfactory tuning parameter values on the fly. Large data sets with ground truth are essential in assessing the performance of a computer vision algorithm. Manually generating document ground truth proved to be very costly and prone to involve subjective errors. We address this problem by using an automatic table ground truth generation system which can efficiently generate a large amount of accurate ground truth suitable for the development of table structure understanding algorithms. This software package is publicly available. The training and testing data set for the algorithm include 1, 125 document pages having 518 table entities and a total of 10, 934 cell entities. The algorithm performed at the 96.76% accuracy rate on the cell level and 98.32% accuracy rate on the table level. We implemented and tested two other published table structure understanding algorithms. In the same data set, with the perfect column structure as the input, the other two algorithms performed at the 88.89% and 82.71% accuracy rate on the table level. Comparing with them, our algorithm demonstrated a favorable result.