Inferring structure information from table images
A computer implemented method includes rendering a document page as an image; detecting tables, columns, and other associated table objects within the image via one or more table recognition models that model objects in the image as overlapping bounding boxes; transforming the set of objects into a structured representation of the table; extracting data from the objects into the structured representation; and exporting the table into the desired output format.
1. A computer implemented method comprising:
detecting a table within a document image;
detecting table objects within the table via a table structure recognition and interpretation model that models table structure and interpretation as a set of overlapping bounding boxes within an image;
transforming the table objects into a structured table representation in part by interpreting overlapping table objects as a hierarchical relationship between table objects;
extracting data from the table objects into the structured table representation; and
exporting the structured table representation and its data into a final output format.
2. The method of claim 1 wherein the table objects comprise columns, rows, and supercells.
3. The method of claim 2 wherein the table objects further comprise column headers, row headers, and subheaders.
4. The method of claim 1 wherein detecting a table within a document comprises providing images of pages of the document as input to a table detection neural network model trained on images of tables with different types of bounding boxes to detect tables.
5. The method of claim 4 wherein the table detection neural network model comprises a convolutional neural network followed by a transformer encoder and transformer decoder.
6. The method of claim 1 wherein the table structure recognition and interpretation model comprises a convolutional neural network followed by a transformer encoder and transformer decoder.
7. The method of claim 6 wherein training data for the table structure recognition and interpretation model comprises images of tables having labels for table objects within the images derived from corresponding structural information for each table.
8. A machine-readable storage device having instructions for execution by a processor of a machine to cause the processor to perform operations to perform a method comprising:
detecting a table within a document image;
detecting table objects within the table via a table structure recognition and interpretation model that models table structure and interpretation as a set of overlapping bounding boxes within an image;
transforming the table objects into a structured table representation;
extracting data from the table objects into the structured table representation; and
exporting the structured table representation and its data into a final output format.
9. The device of claim 8 wherein the table objects comprise columns, rows, and supercells.
10. The device of claim 9 wherein the table objects further comprise column headers, row headers, and subheaders.
11. The device of claim 8 wherein detecting a table within a document comprises providing images of pages of the document as input to a table detection neural network model trained to detect tables.
12. The device of claim 11 wherein the table detection neural network model comprises a convolutional neural network followed by a transformer encoder and transformer decoder.
13. The device of claim 8 wherein the table structure recognition and interpretation model comprises a convolutional neural network followed by a transformer encoder and transformer decoder.
14. The device of claim 13 wherein training data for the table structure recognition and interpretation model comprises images of tables having labels for table objects within the images derived from corresponding structural information for each table.
15. A device comprising:
a processor; and
a memory device coupled to the processor and having a program stored thereon for execution by the processor to perform operations to perform a method comprising:
detecting a table within a document image;
detecting table objects within the table via a table structure recognition and interpretation model that models table structure and interpretation as a set of overlapping bounding boxes within an image;
transforming the table objects into a structured table representation;
extracting data from the table objects into the structured table representation; and
exporting the structured table representation and its data into a final output format.
16. The device of claim 15 wherein the table objects comprise columns, rows, and supercells.
17. The device of claim 16 wherein the table objects further comprise column headers, row headers, and subheaders.
18. The device of claim 15 wherein detecting a table within a document comprises providing images of pages of the document as input to a table detection neural network model trained to detect tables, wherein the table detection neural network model comprises a convolutional neural network followed by a transformer encoder and transformer decoder.
19. The device of claim 15 wherein the table structure recognition and interpretation model comprises a convolutional neural network followed by a transformer encoder and transformer decoder, and wherein training data for the table structure recognition and interpretation model comprises images of tables having labels for table objects within the images derived from corresponding structural information for each table.