Table Recognition Seminar SS21 HPI
Implementing and improving
Koci, E., Thiele, M., Romero, O. and Lehner, W. 2019. A Genetic-Based Search for Adaptive Table Recognition in Spreadsheets. International Conference on Document Analysis and Recognition (ICDAR)
This algorithm uses a graph representation of annotated regions within a spreadsheet to do table recognition. The annotated regions ("label regions") are turned into nodes within a graph. The edges are formed by horizontal or vertical alignment of these label regions.
The components in the graph represent table definitions. All label regions (=nodes) within a component belong to the same table.
By removing edges we can create components. If we iterate over all possible edge combinations we will find all possible
table definitions. This is exactly what the ExhaustiveSearch does.
However, the search space scales with the edge count within the graph, and makes the exhaustive search unfeasible. We
therefore introduce GeneticSearch, a genetic approach to scan a part of the total search space. It is used on graphs
with more than 10 edges.
The quality of a graph with a specific edge list is measured with metrics defined in FitnessRater.
This algorithm will cross validate the machine learning approach to tune the metrics to a specific dataset. Be aware that the computation takes some time and therefore should be run on a server machine.
To run all possible combinations of improvments, datasets and noise and to do three differently seeded runs, simply use
python3 src/run_evaluation.py
To just run specific tests, use
python3 src/run_cross_validation.py --dataset DATASET [--seed SEED] [--noise] [--improvement {NoImprovement,EdgeMutationProbability,EdgeMutationProbabilityExtreme,AvgDegreeCut}]