Techniques for image content extraction

A computer-implemented method, comprising:

• detecting a first object at a first location in input image data with a first machine learning (ML) model, the first object comprising a first table image including a gridded table;
• detecting a second object at a second location in the input image data with the first ML model, the second object comprising a second table image including a non-gridded table;
• detecting a first set of column bounding boxes corresponding to the gridded table in the first table image with a second ML model;
• detecting a second set of column bounding boxes corresponding to the non-gridded table in the second table image with the second ML model;
• identifying first text data including a first set of textual bounding boxes from the first table image and second text data including a second set of textual bounding boxes from the second table image;
• and
• performing a first row extraction process on the first table image based on the first table image including the gridded table and perform a second row extraction process on the second table image based on the second table including the non-gridded table,
• the first row extraction process comprising:
• rotating the first table image to transpose columns and rows in the gridded table to produce a transposed table image,
• detecting a set of row bounding boxes corresponding to the gridded table in the transposed table image with the second ML model, and
• determining first structured data corresponding to the gridded table based on the first set of column bounding boxes, the set of row bounding boxes corresponding to the gridded table, and the first set of textual bounding boxes, and
• the second row extraction process comprising:
• determining a set of rows corresponding to the non-gridded table based on the second text data and the second set of column bounding boxes, and
• determining second structured data corresponding to the non-gridded table based on the second set of column bounding boxes, the set of rows corresponding to the second table, and the second text data.

A computer-program product tangibly embodied in a non-transitory machine-readable storage medium, the computer-program product including instructions operable to cause a processor to perform operations comprising:

• detect a first object at a first location in input image data with a first machine learning (ML) model, the first object comprising a first table image including a gridded table;
• detect a second object at a second location in the input image data with the first ML model, the second object comprising a second table image including a non-gridded table;
• detect a first set of column bounding boxes corresponding to the gridded table in the first table image with a second ML model;
• detect a second set of column bounding boxes corresponding to the non-gridded table in the second table image with the second ML model;
• identify first text data including a first set of textual bounding boxes from the first table image and second text data including a second set of textual bounding boxes from the second table image;
• and
• perform a first row extraction process on the first table image based on the first table image including the gridded table and perform a second row extraction process on the second table image based on the second table including the non-gridded table,
• the first row extraction process comprising:
• rotate the first table image to transpose columns and rows in the gridded table to produce a transposed table image,
• detect a set of row bounding boxes corresponding to the gridded table in the transposed table image with the second ML model, and
• determine first structured data corresponding to the gridded table based on the first set of column bounding boxes, the set of row bounding boxes corresponding to the gridded table, and the first set of textual bounding boxes, and
• the second row extraction process comprising:
• determine a set of rows corresponding to the non-gridded table based on the second text data and the second set of column bounding boxes, and
• determine second structured data corresponding to the non-gridded table based on the second set of column bounding boxes, the set of rows corresponding to the second table, and the second text data.